diff --git "a/data/tasks.jsonl" "b/data/tasks.jsonl"
new file mode 100644--- /dev/null
+++ "b/data/tasks.jsonl"
@@ -0,0 +1,100 @@
+{"context_files": ["task_files/lgr100-001-duplicate-payment-mar/documents/duplicate-payment-control.md"], "gold_output": {"expected_state_assertions": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": "OSINV-5530,OSINV-5555,TPS-11907", "name": "paid_set", "sql": "SELECT GROUP_CONCAT(invoice) FROM (SELECT invoice FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid' ORDER BY invoice)", "type": "sql"}, {"expect": 1, "name": "exactly_one_rejection", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='rejected'", "type": "sql"}, {"expect": "duplicate", "name": "duplicate_rejected_with_code", "sql": "SELECT reason_code FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='5521-OPS'", "type": "sql"}, {"expect": "paid", "name": "lookalike_was_paid", "sql": "SELECT disposition FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='OSINV-5555'", "type": "sql"}, {"expect": 34450.0, "name": "paid_total_rederived", "sql": "SELECT ROUND(total_paid,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": 1, "name": "original_untouched", "sql": "SELECT closed FROM erp_vend_trans WHERE invoice='OSINV-5521'", "type": "sql"}], "submit_answer": {"amount_prevented": 18400.0, "duplicate_found": "yes", "duplicate_invoice": "5521-OPS", "duplicate_of": "OSINV-5521", "run_id": "PR-00001", "total_paid": 34450.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "anomaly_triage", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-001-duplicate-payment-mar", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "docs/HARD-LAYER-DESIGN.md M1/M3/M4/M5/M7; chaos pattern 8 (near-duplicate invoice numbers, ~1.5% duplicate-payment leakage) in research/domain-workflows.md; disbursement controls in research/external/articles/vendor-master-bec-fraud--afp-truist.md", "provenance": "ported", "version": "1.0.0", "walk_len": 11, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 08:05, Friday 6 March**\n\nOffice-supplies run today — vendor group OFFICE, operating account. Cash is fine this week\nso it should be a quick one.\n\nBefore you commit it, run the duplicate screen per SOP-AP-07. We got burned on this in Q4\nand the auditors flagged it, so it's on every run now whether or not cash is tight. Be\ncareful though — last time someone binned a perfectly good invoice because the amount looked\nfamiliar and the vendor was not happy.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PR-00001"], "field": "run_id", "type": "contains_all"}, {"expect": "yes", "field": "duplicate_found", "type": "yes_no"}, {"expect": ["5521-OPS"], "field": "duplicate_invoice", "forbid": ["OSINV-5555"], "type": "contains_all"}, {"expect": ["OSINV-5521"], "field": "duplicate_of", "type": "contains_all"}, {"expect": 34450.0, "field": "total_paid", "gt_sql": "SELECT ROUND(SUM(net_amount),2) FROM erp_payment_run_lines WHERE disposition='paid'", "tol_abs": 0.05, "type": "number"}, {"expect": 18400.0, "field": "amount_prevented", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": "OSINV-5530,OSINV-5555,TPS-11907", "name": "paid_set", "sql": "SELECT GROUP_CONCAT(invoice) FROM (SELECT invoice FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid' ORDER BY invoice)", "type": "sql"}, {"expect": 1, "name": "exactly_one_rejection", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='rejected'", "type": "sql"}, {"expect": "duplicate", "name": "duplicate_rejected_with_code", "sql": "SELECT reason_code FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='5521-OPS'", "type": "sql"}, {"expect": "paid", "name": "lookalike_was_paid", "sql": "SELECT disposition FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='OSINV-5555'", "type": "sql"}, {"expect": 34450.0, "name": "paid_total_rederived", "sql": "SELECT ROUND(total_paid,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": 1, "name": "original_untouched", "sql": "SELECT closed FROM erp_vend_trans WHERE invoice='OSINV-5521'", "type": "sql"}, {"tables": ["answers", "erp_payment_runs", "erp_payment_run_lines", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs", "email"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-001-duplicate-payment-mar", "task_name": "anomaly_triage/duplicate-payment-mar", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"reopened_invoice": "LHINV-001", "returned_amount": 18000.0, "returned_payment_ref": "LHPAY-77", "true_open_balance": 35200.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "bank_rec", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-002-ach-return-mar", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "gap found in repo sweep (research/external/repos/INDEX.md): payment failure after settlement — ACH returns/reversals", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 07:48**\n\nThere's a return on this morning's bank file for Lamna. What does it undo, and what do they actually owe us now? The ERP hasn't caught up yet.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["LHPAY-77"], "field": "returned_payment_ref", "type": "contains_all"}, {"expect": 18000.0, "field": "returned_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["LHINV-001"], "field": "reopened_invoice", "type": "contains_all"}, {"expect": 35200.0, "field": "true_open_balance", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-002-ach-return-mar", "task_name": "bank_rec/ach-return-mar", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"reopened_invoice": "LHINV-001", "returned_amount": 18000.0, "returned_payment_ref": "LHPAY-77", "true_open_balance": 35200.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "bank_rec", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-003-ach-return-mar-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of bank_rec/ach-return-mar via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: gap found in repo sweep (research/external/repos/INDEX.md): payment failure after settlement — ACH returns/reversals", "provenance": "variant", "version": "1.0.0", "walk_len": 5, "walk_servers": ["erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 07:48**\n\nThere's a return on this morning's bank file for Lamna. What does it undo, and what do they actually owe us now? The ERP hasn't caught up yet.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["LHPAY-77"], "field": "returned_payment_ref", "type": "contains_all"}, {"expect": 18000.0, "field": "returned_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["LHINV-001"], "field": "reopened_invoice", "type": "contains_all"}, {"expect": 35200.0, "field": "true_open_balance", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-003-ach-return-mar-esc-burie-quiet", "task_name": "bank_rec/ach-return-mar-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"bank_only_ref": "DEP-9911", "books_only_ref": "PMT-2104", "discrepancy_amount": 45.0, "discrepancy_ref": "PMT-2102", "matched_count": 2}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "bank_rec", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-004-statement-divergence-feb", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#1; grading taxonomy = Microsoft recon-agent buckets (Matched / Discrepancy / bank-only / books-only)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 09:05**\n\nReceipts for 24-28 Feb: bank against books. Anything not lining up? Classify whatever you find.\n", "rubric": {"checks": {"answer_checks": [{"expect": 2, "field": "matched_count", "tol_abs": 0, "type": "number"}, {"expect": ["PMT-2102"], "field": "discrepancy_ref", "type": "contains_all"}, {"expect": 45.0, "field": "discrepancy_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["DEP-9911"], "field": "bank_only_ref", "type": "contains_all"}, {"expect": ["PMT-2104"], "field": "books_only_ref", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-004-statement-divergence-feb", "task_name": "bank_rec/statement-divergence-feb", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"bank_only_ref": "DEP-9911", "books_only_ref": "PMT-2104", "discrepancy_amount": 45.0, "discrepancy_ref": "PMT-2102", "matched_count": 2}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "bank_rec", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-005-statement-divergence-feb-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of bank_rec/statement-divergence-feb via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#1; grading taxonomy = Microsoft recon-agent buckets (Matched / Discrepancy / bank-only / books-only)", "provenance": "variant", "version": "1.0.0", "walk_len": 4, "walk_servers": ["erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 09:05**\n\nReceipts for 24-28 Feb: bank against books. Anything not lining up? Classify whatever you find.\n", "rubric": {"checks": {"answer_checks": [{"expect": 2, "field": "matched_count", "tol_abs": 0, "type": "number"}, {"expect": ["PMT-2102"], "field": "discrepancy_ref", "type": "contains_all"}, {"expect": 45.0, "field": "discrepancy_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["DEP-9911"], "field": "bank_only_ref", "type": "contains_all"}, {"expect": ["PMT-2104"], "field": "books_only_ref", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-005-statement-divergence-feb-esc-burie-quiet", "task_name": "bank_rec/statement-divergence-feb-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-006-brief-caterpillar/documents/template--business-brief.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"equity_fy2024": 19494000000, "internal_ar_relationship": "none", "lt_de_ratio_3yr_avg": 1.43, "lt_debt_fy2024": 27351000000}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "business_brief", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-006-brief-caterpillar", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:business_brief (section schema incl. internal AR/AP fusion) converted from LLM-judge rubric to structured fields; ratios from user-brief Caterpillar example", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "erp", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email, Tue 07:55**\n\nI'm presenting the Caterpillar credit brief Thursday. Can you do the leverage section — the debt and equity figures, the three-year ratio picture — and check whether we already trade with them?\n", "rubric": {"checks": {"answer_checks": [{"expect": 27351000000, "field": "lt_debt_fy2024", "tol_rel": 0.001, "type": "number"}, {"expect": 19494000000, "field": "equity_fy2024", "tol_rel": 0.001, "type": "number"}, {"expect": 1.4253, "field": "lt_de_ratio_3yr_avg", "tol_abs": 0.02, "type": "number"}, {"field": "internal_ar_relationship", "type": "none_answer"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "erp"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-006-brief-caterpillar", "task_name": "business_brief/brief-caterpillar", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-007-brief-caterpillar-esc-burie-quiet/documents/template--business-brief.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"equity_fy2024": 19494000000, "internal_ar_relationship": "none", "lt_de_ratio_3yr_avg": 1.43, "lt_debt_fy2024": 27351000000}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "business_brief", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-007-brief-caterpillar-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of business_brief/brief-caterpillar via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark:business_brief (section schema incl. internal AR/AP fusion) converted from LLM-judge rubric to structured fields; ratios from user-brief Caterpillar example", "provenance": "variant", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "erp", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email, Tue 07:55**\n\nI'm presenting the Caterpillar credit brief Thursday. Can you do the leverage section — the debt and equity figures, the three-year ratio picture — and check whether we already trade with them?\n", "rubric": {"checks": {"answer_checks": [{"expect": 27351000000, "field": "lt_debt_fy2024", "tol_rel": 0.001, "type": "number"}, {"expect": 19494000000, "field": "equity_fy2024", "tol_rel": 0.001, "type": "number"}, {"expect": 1.4253, "field": "lt_de_ratio_3yr_avg", "tol_abs": 0.02, "type": "number"}, {"field": "internal_ar_relationship", "type": "none_answer"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "erp"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-007-brief-caterpillar-esc-burie-quiet", "task_name": "business_brief/brief-caterpillar-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-008-brief-caterpillar-v2/documents/policy--counterparty-credit.md", "task_files/lgr100-008-brief-caterpillar-v2/documents/template--business-brief.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"internal_relationship": "none", "leverage_classification": "moderate", "lt_de_2022": 1.62, "lt_de_2023": 1.25, "lt_de_2024": 1.4, "lt_de_ratio_3yr_avg": 1.43}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "harder", "family": "business_brief", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-008-brief-caterpillar-v2", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "grow: escalated from business_brief/brief-caterpillar (haiku passed 2/2 -> too_easy). Escalations: per-year ratio fields, leverage classification requires reading the credit policy bands, walk 8->10.", "provenance": "ported", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "erp", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email, Tue 07:55**\n\nCaterpillar credit brief for Thursday. I need the leverage worked per year rather than just the average, classified the way our policy wants it, and the usual check on whether we already trade with them. Follow the brief template.\n", "rubric": {"checks": {"answer_checks": [{"expect": 1.618, "field": "lt_de_2022", "tol_abs": 0.02, "type": "number"}, {"expect": 1.255, "field": "lt_de_2023", "tol_abs": 0.02, "type": "number"}, {"expect": 1.403, "field": "lt_de_2024", "tol_abs": 0.02, "type": "number"}, {"expect": 1.425, "field": "lt_de_ratio_3yr_avg", "tol_abs": 0.02, "type": "number"}, {"expect": ["moderate"], "field": "leverage_classification", "type": "contains_all"}, {"field": "internal_relationship", "type": "none_answer"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-008-brief-caterpillar-v2", "task_name": "business_brief/brief-caterpillar-v2", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"internal_ar_relationship": "none", "internal_open_ar_usd": 0, "lt_de_ratio_3yr_avg": 0.0326, "net_income_fy2025": 4335000000.0, "revenue_fy2025": 34639000000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "business_brief_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-009-brief-advanced-micro-devices-inc", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark business_brief item business_brief-007 (Advanced Micro Devices Inc); the JUDGEMENT is ported and FB's DSPy LLM-judge prose scoring is dropped per docs/INGESTION.md judgement_port. Figures recomputed from the frozen SEC XBRL snapshot; the internal_ar_relationship field is this world's addition — it is answerable only from the ERP, not the filings (scale/profitability @ 2025-12-27 (10-K); 3yr avg long-term D/E; no internal exposure (trap))", "provenance": "ported", "version": "1.0.0", "walk_len": 7, "walk_servers": ["erp", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email**\n\nPutting together the counterparty brief on Advanced Micro Devices Inc. I need the public numbers — scale, profitability and the leverage picture — and, importantly, whether we already have any exposure to them on our own books. Don't assume either way, check.\n", "rubric": {"checks": {"answer_checks": [{"expect": 34639000000.0, "field": "revenue_fy2025", "tol_rel": 0.001, "type": "number"}, {"expect": 4335000000.0, "field": "net_income_fy2025", "tol_rel": 0.001, "type": "number"}, {"expect": 0.0326, "field": "lt_de_ratio_3yr_avg", "tol_abs": 0.02, "type": "number"}, {"field": "internal_ar_relationship", "type": "none_answer"}, {"expect": 0, "field": "internal_open_ar_usd", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "erp"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-009-brief-advanced-micro-devices-inc", "task_name": "business_brief_fb/brief-advanced-micro-devices-inc", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"internal_ar_relationship": "BRF-20", "internal_open_ar_usd": 42132, "lt_de_ratio_3yr_avg": -6.1185, "net_income_fy2025": 111000000.0, "revenue_fy2025": 54633000000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "business_brief_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-010-brief-american-airlines-group", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark business_brief item business_brief-158 (American Airlines Group); the JUDGEMENT is ported and FB's DSPy LLM-judge prose scoring is dropped per docs/INGESTION.md judgement_port. Figures recomputed from the frozen SEC XBRL snapshot; the internal_ar_relationship field is this world's addition — it is answerable only from the ERP, not the filings (scale/profitability @ 2025-12-31 (10-K); 3yr avg long-term D/E; internal exposure BRF-20 = 42132)", "provenance": "ported", "version": "1.0.0", "walk_len": 7, "walk_servers": ["erp", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email**\n\nPutting together the counterparty brief on American Airlines Group. I need the public numbers — scale, profitability and the leverage picture — and, importantly, whether we already have any exposure to them on our own books. Don't assume either way, check.\n", "rubric": {"checks": {"answer_checks": [{"expect": 54633000000.0, "field": "revenue_fy2025", "tol_rel": 0.001, "type": "number"}, {"expect": 111000000.0, "field": "net_income_fy2025", "tol_rel": 0.001, "type": "number"}, {"expect": -6.1185, "field": "lt_de_ratio_3yr_avg", "tol_abs": 0.02, "type": "number"}, {"expect": ["BRF-20"], "field": "internal_ar_relationship", "type": "contains_all"}, {"expect": 42132, "field": "internal_open_ar_usd", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "erp"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-010-brief-american-airlines-group", "task_name": "business_brief_fb/brief-american-airlines-group", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"internal_ar_relationship": "BRF-00", "internal_open_ar_usd": 41732, "lt_de_ratio_3yr_avg": 1.3671, "net_income_fy2025": 112010000000.0, "revenue_fy2025": 416161000000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "business_brief_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-011-brief-apple-inc", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark business_brief item business_brief-006 (Apple Inc); the JUDGEMENT is ported and FB's DSPy LLM-judge prose scoring is dropped per docs/INGESTION.md judgement_port. Figures recomputed from the frozen SEC XBRL snapshot; the internal_ar_relationship field is this world's addition — it is answerable only from the ERP, not the filings (scale/profitability @ 2025-09-27 (10-K); 3yr avg long-term D/E; internal exposure BRF-00 = 41732)", "provenance": "ported", "version": "1.0.0", "walk_len": 7, "walk_servers": ["erp", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email**\n\nPutting together the counterparty brief on Apple Inc. I need the public numbers — scale, profitability and the leverage picture — and, importantly, whether we already have any exposure to them on our own books. Don't assume either way, check.\n", "rubric": {"checks": {"answer_checks": [{"expect": 416161000000.0, "field": "revenue_fy2025", "tol_rel": 0.001, "type": "number"}, {"expect": 112010000000.0, "field": "net_income_fy2025", "tol_rel": 0.001, "type": "number"}, {"expect": 1.3671, "field": "lt_de_ratio_3yr_avg", "tol_abs": 0.02, "type": "number"}, {"expect": ["BRF-00"], "field": "internal_ar_relationship", "type": "contains_all"}, {"expect": 41732, "field": "internal_open_ar_usd", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "erp"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-011-brief-apple-inc", "task_name": "business_brief_fb/brief-apple-inc", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-012-deduction-coding-mar/documents/copper-ridge-po.md", "task_files/lgr100-012-deduction-coding-mar/documents/deduction-policy.md", "task_files/lgr100-012-deduction-coding-mar/documents/sable-finch-allowance.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"chargeback_amount": 5400.0, "cinv701_reason": "pricing_variance", "cinv702_reason": "promotional_allowance", "cinv703_reason": "write_off_immaterial", "cinv704_reason": "unauthorized", "conceded_amount": 1950.0, "total_deductions": 7350.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cash_app", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-012-deduction-coding-mar", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "closes tracked coverage gaps art.deduction_coding and fb.erp.deductions; deduction taxonomy and routing from research/external/articles/cash-application--{highradius,stuut,zamp}.md; docs/HARD-LAYER-DESIGN.md M3/M4/M5", "provenance": "ported", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Ana Duarte · AR Manager · Teams 08:45**\n\nFour short-pays came in this week — CINV-701 through CINV-704. Can you code and route them\nper SOP-AR-06 before the AR bridge goes out?\n\nDon't just take the reason off the remittance. We had one last quarter coded as a damages\nclaim on the customer's say-so and it turned out to be a price they'd never agreed; we\nconceded 12k we should have charged back.\n\nConceded and chargeback go on separate lines of the bridge — don't net them.\n", "rubric": {"checks": {"answer_checks": [{"expect": 7350.0, "field": "total_deductions", "gt_sql": "SELECT ROUND(SUM(amount-settled),2) FROM erp_cust_trans WHERE invoice LIKE 'CINV-70%'", "tol_abs": 0.05, "type": "number"}, {"expect": ["pricing_variance"], "field": "cinv701_reason", "forbid": ["shortage_damage", "unauthorized", "promotional_allowance", "write_off"], "type": "contains_all"}, {"expect": ["promotional_allowance"], "field": "cinv702_reason", "forbid": ["unauthorized", "pricing_variance", "shortage_damage"], "type": "contains_all"}, {"expect": ["write_off_immaterial"], "field": "cinv703_reason", "forbid": ["shortage_damage", "pricing_variance", "unauthorized"], "type": "contains_all"}, {"expect": ["unauthorized"], "field": "cinv704_reason", "forbid": ["shortage_damage", "pricing_variance", "promotional_allowance"], "type": "contains_all"}, {"expect": 1950.0, "field": "conceded_amount", "tol_abs": 0.05, "type": "number"}, {"expect": 5400.0, "field": "chargeback_amount", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs", "email"], "type": "required_servers"}, {"n": 2, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-012-deduction-coding-mar", "task_name": "cash_app/deduction-coding-mar", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"dep501_invoices": "LHINV-001, LHINV-002", "dep502_invoices": "LHINV-003", "dep503_invoices": "none", "unapplied_amount": 4000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "cash_app", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-013-remittance-batch-mar02", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#4 (HighRadius/Auditoria pattern: remittance emails + bank export -> apply); 46% unapplied-cash stat", "provenance": "ported", "version": "1.0.0", "walk_len": 7, "walk_servers": ["email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 08:12**\n\nThree deposits from Lamna in this morning's lockbox. Can you work out how they apply? Don't guess on anything you can't tie out.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["LHINV-001", "LHINV-002"], "field": "dep501_invoices", "type": "contains_all"}, {"expect": ["LHINV-003"], "field": "dep502_invoices", "type": "contains_all"}, {"field": "dep503_invoices", "forbid": ["LHINV-001", "LHINV-002", "LHINV-003"], "type": "none_answer"}, {"expect": 4000.0, "field": "unapplied_amount", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets", "email"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-013-remittance-batch-mar02", "task_name": "cash_app/remittance-batch-mar02", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"dep501_invoices": "LHINV-001, LHINV-002", "dep502_invoices": "LHINV-003", "dep503_invoices": "none", "unapplied_amount": 4000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "cash_app", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-014-remittance-batch-mar02-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of cash_app/remittance-batch-mar02 via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#4 (HighRadius/Auditoria pattern: remittance emails + bank export -> apply); 46% unapplied-cash stat", "provenance": "variant", "version": "1.0.0", "walk_len": 7, "walk_servers": ["email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 08:12**\n\nThree deposits from Lamna in this morning's lockbox. Can you work out how they apply? Don't guess on anything you can't tie out.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["LHINV-001", "LHINV-002"], "field": "dep501_invoices", "type": "contains_all"}, {"expect": ["LHINV-003"], "field": "dep502_invoices", "type": "contains_all"}, {"field": "dep503_invoices", "forbid": ["LHINV-001", "LHINV-002", "LHINV-003"], "type": "none_answer"}, {"expect": 4000.0, "field": "unapplied_amount", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets", "email"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-014-remittance-batch-mar02-esc-burie-quiet", "task_name": "cash_app/remittance-batch-mar02-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-015-cesp-four-week/documents/sop--cash-forecast-assumptions.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"four_week_net": -322000.0, "week1_net": 14000.0, "week2_net": -169000.0, "week3_net": 51000.0, "week4_net": -218000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cash_forecast", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-015-cesp-four-week", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#8 (13-week direct method, scoped to 4 weeks); assumptions doc = consult-don't-know", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email, Mon 07:30**\n\nBoard call Thursday — I need CESP's cash position for the next four weeks, week by week starting today. Use our standard assumptions.\n", "rubric": {"checks": {"answer_checks": [{"expect": 14000.0, "field": "week1_net", "tol_abs": 0.05, "type": "number"}, {"expect": -169000.0, "field": "week2_net", "tol_abs": 0.05, "type": "number"}, {"expect": 51000.0, "field": "week3_net", "tol_abs": 0.05, "type": "number"}, {"expect": -218000.0, "field": "week4_net", "tol_abs": 0.05, "type": "number"}, {"expect": -322000.0, "field": "four_week_net", "tol_abs": 0.1, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-015-cesp-four-week", "task_name": "cash_forecast/cesp-four-week", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-016-cesp-four-week-esc-burie-quiet/documents/sop--cash-forecast-assumptions.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"four_week_net": -322000.0, "week1_net": 14000.0, "week2_net": -169000.0, "week3_net": 51000.0, "week4_net": -218000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cash_forecast", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-016-cesp-four-week-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of cash_forecast/cesp-four-week via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#8 (13-week direct method, scoped to 4 weeks); assumptions doc = consult-don't-know", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · email, Mon 07:30**\n\nBoard call Thursday — I need CESP's cash position for the next four weeks, week by week starting today. Use our standard assumptions.\n", "rubric": {"checks": {"answer_checks": [{"expect": 14000.0, "field": "week1_net", "tol_abs": 0.05, "type": "number"}, {"expect": -169000.0, "field": "week2_net", "tol_abs": 0.05, "type": "number"}, {"expect": 51000.0, "field": "week3_net", "tol_abs": 0.05, "type": "number"}, {"expect": -218000.0, "field": "week4_net", "tol_abs": 0.05, "type": "number"}, {"expect": -322000.0, "field": "four_week_net", "tol_abs": 0.1, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-016-cesp-four-week-esc-burie-quiet", "task_name": "cash_forecast/cesp-four-week-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-017-period-lock-correction/documents/period-close-policy.md"], "gold_output": {"expected_state_assertions": [{"expect": "posted", "name": "correction_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": "2026-03", "name": "booked_to_open_period", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": 1, "name": "nothing_posted_into_january", "sql": "SELECT COUNT(*) FROM erp_ledger_journals WHERE period_id='2026-01'", "type": "sql"}, {"expect": "600200", "name": "debit_to_software", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND debit>0", "type": "sql"}, {"expect": "600300", "name": "credit_reverses_te", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND credit>0", "type": "sql"}, {"expect": 0.0, "name": "reclass_nets_to_zero", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "tol_abs": 0.005, "type": "sql"}, {"expect": "posted", "name": "original_untouched", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00900'", "type": "sql"}], "submit_answer": {"amount": 8400.0, "january_status": "closed", "journal_id": "GJ-00002", "journal_state": "posted", "posting_period": "2026-03"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "close_mgmt", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-017-period-lock-correction", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "closes tracked coverage gap erpnext.period_close_lock (sim/coverage.py); period-status model from research/odoo-domain.md (Odoo company lock dates) and research/erp-mcp-tool-census.md (mcp-erp close_period/reopen_period); docs/HARD-LAYER-DESIGN.md M1/M3/M5", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Controller · Teams 09:12**\n\nAudit picked up a coding error from January — GJ-00900, the Northwind analytics\nsubscription, $8,400 sitting in travel & entertainment instead of software. It's a straight\nreclass between expense accounts, no cash impact.\n\nCan you book the correction and tell me where it landed? January's numbers have already gone\nout, so be careful about the dating — check the calendar before you date it, ours doesn't\ntrack the calendar month.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["GJ-00002"], "field": "journal_id", "type": "contains_all"}, {"expect": ["2026-03"], "field": "posting_period", "type": "contains_all"}, {"expect": ["closed"], "field": "january_status", "type": "contains_all"}, {"expect": ["posted"], "field": "journal_state", "type": "contains_all"}, {"expect": 8400.0, "field": "amount", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": "posted", "name": "correction_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": "2026-03", "name": "booked_to_open_period", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": 1, "name": "nothing_posted_into_january", "sql": "SELECT COUNT(*) FROM erp_ledger_journals WHERE period_id='2026-01'", "type": "sql"}, {"expect": "600200", "name": "debit_to_software", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND debit>0", "type": "sql"}, {"expect": "600300", "name": "credit_reverses_te", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND credit>0", "type": "sql"}, {"expect": 0.0, "name": "reclass_nets_to_zero", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "tol_abs": 0.005, "type": "sql"}, {"expect": "posted", "name": "original_untouched", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00900'", "type": "sql"}, {"tables": ["answers", "erp_ledger_journals", "erp_ledger_journal_lines", "erp_approval_requests", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-017-period-lock-correction", "task_name": "close_mgmt/period-lock-correction", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-018-period-lock-correction-esc-burie-quiet/documents/period-close-policy.md"], "gold_output": {"expected_state_assertions": [{"expect": "posted", "name": "correction_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": "2026-03", "name": "booked_to_open_period", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": 1, "name": "nothing_posted_into_january", "sql": "SELECT COUNT(*) FROM erp_ledger_journals WHERE period_id='2026-01'", "type": "sql"}, {"expect": "600200", "name": "debit_to_software", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND debit>0", "type": "sql"}, {"expect": "600300", "name": "credit_reverses_te", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND credit>0", "type": "sql"}, {"expect": 0.0, "name": "reclass_nets_to_zero", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "tol_abs": 0.005, "type": "sql"}, {"expect": "posted", "name": "original_untouched", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00900'", "type": "sql"}], "submit_answer": {"amount": 8400.0, "january_status": "closed", "journal_id": "GJ-00002", "journal_state": "posted", "posting_period": "2026-03"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "close_mgmt", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-018-period-lock-correction-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of close_mgmt/period-lock-correction via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: closes tracked coverage gap erpnext.period_close_lock (sim/coverage.py); period-status model from research/odoo-domain.md (Odoo company lock dates) and research/erp-mcp-tool-census.md (mcp-erp close_period/reopen_period); docs/HARD-LAYER-DESIGN.md M1/M3/M5", "provenance": "variant", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Controller · Teams 09:12**\n\nAudit picked up a coding error from January — GJ-00900, the Northwind analytics\nsubscription, $8,400 sitting in travel & entertainment instead of software. It's a straight\nreclass between expense accounts, no cash impact.\n\nCan you book the correction and tell me where it landed? January's numbers have already gone\nout, so be careful about the dating — check the calendar before you date it, ours doesn't\ntrack the calendar month.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["GJ-00002"], "field": "journal_id", "type": "contains_all"}, {"expect": ["2026-03"], "field": "posting_period", "type": "contains_all"}, {"expect": ["closed"], "field": "january_status", "type": "contains_all"}, {"expect": ["posted"], "field": "journal_state", "type": "contains_all"}, {"expect": 8400.0, "field": "amount", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": "posted", "name": "correction_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": "2026-03", "name": "booked_to_open_period", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "type": "sql"}, {"expect": 1, "name": "nothing_posted_into_january", "sql": "SELECT COUNT(*) FROM erp_ledger_journals WHERE period_id='2026-01'", "type": "sql"}, {"expect": "600200", "name": "debit_to_software", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND debit>0", "type": "sql"}, {"expect": "600300", "name": "credit_reverses_te", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00002' AND credit>0", "type": "sql"}, {"expect": 0.0, "name": "reclass_nets_to_zero", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00002'", "tol_abs": 0.005, "type": "sql"}, {"expect": "posted", "name": "original_untouched", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00900'", "type": "sql"}, {"tables": ["answers", "erp_ledger_journals", "erp_ledger_journal_lines", "erp_approval_requests", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-018-period-lock-correction-esc-burie-quiet", "task_name": "close_mgmt/period-lock-correction-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-019-revenue-recognition-tieout/documents/policy--revenue-recognition-tieout.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"amount_exceptions": "CTR-2047, CTR-2052", "flagged_contracts": "CTR-2047, CTR-2052, CTR-2055, CTR-2058", "over_recognised_usd": 7500.0, "timing_exceptions": "CTR-2055, CTR-2058", "under_recognised_usd": 8000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "close_mgmt", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-019-revenue-recognition-tieout", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-revenue-reconciliation (contracts.xlsx + revenue_schedule.xlsx, flag contracts where recognised revenue misses contract value beyond tolerance OR the schedule does not align with the contract term); the source's LLM-judged RocketChat notification and file-writing checkpoints are dropped for deterministic answer checks, and its single flagged list is split into the two finding classes its own task.md names; docs/HARD-LAYER-DESIGN.md M3 (non-collapse) / M4 (recomputed from the schedule, never read off the register's cached column) / M5 (tolerance and term-alignment rule live only in RAP-04); source checkpoint 1 (asserting the ownCloud path appears in the trajectory) is also dropped as environment-specific - the equivalent here is the required_servers/reads_before_submit trace check", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Iris Bhatt · Revenue Accounting Manager · Teams 07:50**\n\nThe auditors moved the revenue walkthrough to Thursday and the FY26 contract register goes out\nwith the PBC pack. Can you tie it out to the recognition schedule per RAP-04 before I send it?\n\nLast year we handed them a register that footed perfectly and still took two cut-off findings.\nNobody caught them in-house, because the totals agreed. So please don't only foot it.\n\nAlso — don't lean on the recognised column in the register. Someone refreshes that by hand and\nI have no idea when it was last done.\n\nOver- and under-recognised go on separate lines. Netting them is how a real exposure turns into\na rounding difference on the memo, and the auditors ask for the directions anyway.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["CTR-2047", "CTR-2052", "CTR-2055", "CTR-2058"], "field": "flagged_contracts", "forbid": ["CTR-2041", "CTR-2044", "CTR-2061"], "type": "contains_all"}, {"expect": ["CTR-2047", "CTR-2052"], "field": "amount_exceptions", "forbid": ["CTR-2041", "CTR-2044", "CTR-2055", "CTR-2058", "CTR-2061"], "type": "contains_all"}, {"expect": ["CTR-2055", "CTR-2058"], "field": "timing_exceptions", "forbid": ["CTR-2041", "CTR-2044", "CTR-2047", "CTR-2052", "CTR-2061"], "type": "contains_all"}, {"expect": 7500.0, "field": "over_recognised_usd", "gt_sql": "WITH c AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS val FROM sheet_rows WHERE file='contract-register-fy26.xlsx' AND row_no>1), s AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS amt FROM sheet_rows WHERE file='revenue-schedule-fy26.xlsx' AND row_no>1), v AS (SELECT c.cid AS cid, c.val AS val, COALESCE((SELECT SUM(amt) FROM s WHERE s.cid=c.cid),0) AS rec FROM c) SELECT ROUND(COALESCE(SUM(rec-val),0),2) FROM v WHERE rec-val > (CASE WHEN val*0.001>100 THEN val*0.001 ELSE 100 END)", "tol_abs": 0.05, "type": "number"}, {"expect": 8000.0, "field": "under_recognised_usd", "gt_sql": "WITH c AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS val FROM sheet_rows WHERE file='contract-register-fy26.xlsx' AND row_no>1), s AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS amt FROM sheet_rows WHERE file='revenue-schedule-fy26.xlsx' AND row_no>1), v AS (SELECT c.cid AS cid, c.val AS val, COALESCE((SELECT SUM(amt) FROM s WHERE s.cid=c.cid),0) AS rec FROM c) SELECT ROUND(COALESCE(SUM(val-rec),0),2) FROM v WHERE val-rec > (CASE WHEN val*0.001>100 THEN val*0.001 ELSE 100 END)", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-019-revenue-recognition-tieout", "task_name": "close_mgmt/revenue-recognition-tieout", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-020-revenue-recognition-tieout-esc-burie-quiet/documents/policy--revenue-recognition-tieout.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"amount_exceptions": "CTR-2047, CTR-2052", "flagged_contracts": "CTR-2047, CTR-2052, CTR-2055, CTR-2058", "over_recognised_usd": 7500.0, "timing_exceptions": "CTR-2055, CTR-2058", "under_recognised_usd": 8000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "close_mgmt", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-020-revenue-recognition-tieout-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of close_mgmt/revenue-recognition-tieout via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-revenue-reconciliation (contracts.xlsx + revenue_schedule.xlsx, flag contracts where recognised revenue misses contract value beyond tolerance OR the schedule does not align with the contract term); the source's LLM-judged RocketChat notification and file-writing checkpoints are dropped for deterministic answer checks, and its single flagged list is split into the two finding classes its own task.md names; docs/HARD-LAYER-DESIGN.md M3 (non-collapse) / M4 (recomputed from the schedule, never read off the register's cached column) / M5 (tolerance and term-alignment rule live only in RAP-04); source checkpoint 1 (asserting the ownCloud path appears in the trajectory) is also dropped as environment-specific - the equivalent here is the required_servers/reads_before_submit trace check", "provenance": "variant", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Iris Bhatt · Revenue Accounting Manager · Teams 07:50**\n\nThe auditors moved the revenue walkthrough to Thursday and the FY26 contract register goes out\nwith the PBC pack. Can you tie it out to the recognition schedule per RAP-04 before I send it?\n\nLast year we handed them a register that footed perfectly and still took two cut-off findings.\nNobody caught them in-house, because the totals agreed. So please don't only foot it.\n\nAlso — don't lean on the recognised column in the register. Someone refreshes that by hand and\nI have no idea when it was last done.\n\nOver- and under-recognised go on separate lines. Netting them is how a real exposure turns into\na rounding difference on the memo, and the auditors ask for the directions anyway.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["CTR-2047", "CTR-2052", "CTR-2055", "CTR-2058"], "field": "flagged_contracts", "forbid": ["CTR-2041", "CTR-2044", "CTR-2061"], "type": "contains_all"}, {"expect": ["CTR-2047", "CTR-2052"], "field": "amount_exceptions", "forbid": ["CTR-2041", "CTR-2044", "CTR-2055", "CTR-2058", "CTR-2061"], "type": "contains_all"}, {"expect": ["CTR-2055", "CTR-2058"], "field": "timing_exceptions", "forbid": ["CTR-2041", "CTR-2044", "CTR-2047", "CTR-2052", "CTR-2061"], "type": "contains_all"}, {"expect": 7500.0, "field": "over_recognised_usd", "gt_sql": "WITH c AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS val FROM sheet_rows WHERE file='contract-register-fy26.xlsx' AND row_no>1), s AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS amt FROM sheet_rows WHERE file='revenue-schedule-fy26.xlsx' AND row_no>1), v AS (SELECT c.cid AS cid, c.val AS val, COALESCE((SELECT SUM(amt) FROM s WHERE s.cid=c.cid),0) AS rec FROM c) SELECT ROUND(COALESCE(SUM(rec-val),0),2) FROM v WHERE rec-val > (CASE WHEN val*0.001>100 THEN val*0.001 ELSE 100 END)", "tol_abs": 0.05, "type": "number"}, {"expect": 8000.0, "field": "under_recognised_usd", "gt_sql": "WITH c AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS val FROM sheet_rows WHERE file='contract-register-fy26.xlsx' AND row_no>1), s AS (SELECT json_extract(cells,'$[0]') AS cid, CAST(json_extract(cells,'$[2]') AS REAL) AS amt FROM sheet_rows WHERE file='revenue-schedule-fy26.xlsx' AND row_no>1), v AS (SELECT c.cid AS cid, c.val AS val, COALESCE((SELECT SUM(amt) FROM s WHERE s.cid=c.cid),0) AS rec FROM c) SELECT ROUND(COALESCE(SUM(val-rec),0),2) FROM v WHERE val-rec > (CASE WHEN val*0.001>100 THEN val*0.001 ELSE 100 END)", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-020-revenue-recognition-tieout-esc-burie-quiet", "task_name": "close_mgmt/revenue-recognition-tieout-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-021-subledger-tieout-feb/documents/sop--month-end-close-checklist.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"blocking_task_owner": "Jordan Blake", "erp_ap_balance": 80500.0, "erp_ar_balance": 128500.0, "unreconciled_subledger": "AP", "variance_amount": 2500.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "close_mgmt", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-021-subledger-tieout-feb", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#2 (close checklist + subledger tie-out); chaos: workbook vs ERP divergence", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 18:20**\n\nCESP February close is stuck. Which subledger isn't tying, by how much, and who owns that task?\n", "rubric": {"checks": {"answer_checks": [{"expect": 128500.0, "field": "erp_ar_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 80500.0, "field": "erp_ap_balance", "tol_abs": 0.02, "type": "number"}, {"expect": ["AP"], "field": "unreconciled_subledger", "type": "contains_all"}, {"expect": 2500.0, "field": "variance_amount", "tol_abs": 0.02, "type": "number"}, {"expect": ["jordan"], "field": "blocking_task_owner", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-021-subledger-tieout-feb", "task_name": "close_mgmt/subledger-tieout-feb", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-022-subledger-tieout-feb-esc-burie-quiet/documents/sop--month-end-close-checklist.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"blocking_task_owner": "Jordan Blake", "erp_ap_balance": 80500.0, "erp_ar_balance": 128500.0, "unreconciled_subledger": "AP", "variance_amount": 2500.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "close_mgmt", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-022-subledger-tieout-feb-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of close_mgmt/subledger-tieout-feb via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#2 (close checklist + subledger tie-out); chaos: workbook vs ERP divergence", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 18:20**\n\nCESP February close is stuck. Which subledger isn't tying, by how much, and who owns that task?\n", "rubric": {"checks": {"answer_checks": [{"expect": 128500.0, "field": "erp_ar_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 80500.0, "field": "erp_ap_balance", "tol_abs": 0.02, "type": "number"}, {"expect": ["AP"], "field": "unreconciled_subledger", "type": "contains_all"}, {"expect": 2500.0, "field": "variance_amount", "tol_abs": 0.02, "type": "number"}, {"expect": ["jordan"], "field": "blocking_task_owner", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-022-subledger-tieout-feb-esc-burie-quiet", "task_name": "close_mgmt/subledger-tieout-feb-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-023-escalate-sparrow-letter3/documents/policy--collections-dunning-runbook.md"], "gold_output": {"expected_state_assertions": [{"expect": 1, "name": "letter3_posted_today", "sql": "SELECT COUNT(*) FROM erp_collection_letters WHERE account='US-008' AND letter_code='3' AND letter_date='2026-03-02' AND status='Sent'", "type": "row_count"}, {"expect": 3, "name": "no_duplicate_letters", "sql": "SELECT COUNT(*) FROM erp_collection_letters WHERE account='US-008'", "type": "row_count"}, {"expect": 40.0, "name": "letter3_fee", "sql": "SELECT fee FROM erp_collection_letters WHERE account='US-008' AND letter_code='3'", "tol_abs": 0.01, "type": "cell_equals"}, {"expect": "Open", "name": "no_credit_hold", "sql": "SELECT on_hold FROM erp_customers WHERE account='US-008'", "type": "cell_equals"}], "submit_answer": {"credit_hold_placed": "no", "letter_fee": 40.0, "letter_issued": "3", "reasoning": "Letter 2 was sent 14 days ago and SPINV-002 remains unpaid, so letter 3 posts with its $40 fee; the oldest past-due item is at 47 days, below the ~75-day credit-hold threshold, so no hold yet."}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "collections_ops", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-023-escalate-sparrow-letter3", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "wave-2 writes on the D365 hold/letter machinery (research/finance-agent-workflows.md wf#10); dunning ladder AR-SOP-003", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 09:00**\n\nSparrow Retail's review is due today. Take whatever action the runbook calls for at their position — nothing more — and tell me what you did and why.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["3"], "field": "letter_issued", "type": "contains_all"}, {"expect": 40.0, "field": "letter_fee", "tol_abs": 0.01, "type": "number"}, {"expect": "no", "field": "credit_hold_placed", "type": "yes_no"}, {"expect": ["75"], "field": "reasoning", "type": "contains_all"}], "state_checks": [{"tables": ["answers", "erp_collection_letters"], "type": "writes_only"}, {"expect": 1, "name": "letter3_posted_today", "sql": "SELECT COUNT(*) FROM erp_collection_letters WHERE account='US-008' AND letter_code='3' AND letter_date='2026-03-02' AND status='Sent'", "type": "row_count"}, {"expect": 3, "name": "no_duplicate_letters", "sql": "SELECT COUNT(*) FROM erp_collection_letters WHERE account='US-008'", "type": "row_count"}, {"expect": 40.0, "name": "letter3_fee", "sql": "SELECT fee FROM erp_collection_letters WHERE account='US-008' AND letter_code='3'", "tol_abs": 0.01, "type": "cell_equals"}, {"expect": "Open", "name": "no_credit_hold", "sql": "SELECT on_hold FROM erp_customers WHERE account='US-008'", "type": "cell_equals"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-023-escalate-sparrow-letter3", "task_name": "collections_ops/escalate-sparrow-letter3", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"evidence_source": "email attachment (shared AP/AR mailbox), corroborated by the manual invoice log spreadsheet", "found_in_erp": "no", "invoice_number": "INV-2026-047", "outstanding_amount": 18750.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-024-email-invoice-meadow", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "novel: chaos patterns 'invoices existing only in email' (68% manual keying) + 'Excel-only manual log' (research/domain-workflows.md §3)", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 08:05**\n\nDid we ever invoice Meadow Analytics for the February executive workshop? They're a new client and I can't find them. What's outstanding from them?\n", "rubric": {"checks": {"answer_checks": [{"expect": "no", "field": "found_in_erp", "type": "yes_no"}, {"expect": ["INV-2026-047"], "field": "invoice_number", "type": "contains_all"}, {"expect": 18750.0, "field": "outstanding_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["mail"], "field": "evidence_source", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "email"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-024-email-invoice-meadow", "task_name": "cross_system/email-invoice-meadow", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"evidence_source": "email attachment (shared AP/AR mailbox), corroborated by the manual invoice log spreadsheet", "found_in_erp": "no", "invoice_number": "INV-2026-047", "outstanding_amount": 18750.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-025-email-invoice-meadow-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of cross_system/email-invoice-meadow via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: novel: chaos patterns 'invoices existing only in email' (68% manual keying) + 'Excel-only manual log' (research/domain-workflows.md §3)", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 08:05**\n\nDid we ever invoice Meadow Analytics for the February executive workshop? They're a new client and I can't find them. What's outstanding from them?\n", "rubric": {"checks": {"answer_checks": [{"expect": "no", "field": "found_in_erp", "type": "yes_no"}, {"expect": ["INV-2026-047"], "field": "invoice_number", "type": "contains_all"}, {"expect": 18750.0, "field": "outstanding_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["mail"], "field": "evidence_source", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "email"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-025-email-invoice-meadow-esc-burie-quiet", "task_name": "cross_system/email-invoice-meadow-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"difference": 22500.0, "difference_cause": "ICINV-03 ($22,500) — received by the subsidiary 2026-03-01 but not yet booked (in transit)", "parent_ar_balance": 76500.0, "sub_recorded_balance": 54000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-026-intercompany-tieout-feb", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#7; chaos primitive: system A learned the truth before system B", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 17:02**\n\nIntercompany with CES Direct for February. What do we say they owe, what do they say, and what's behind any difference?\n", "rubric": {"checks": {"answer_checks": [{"expect": 76500.0, "field": "parent_ar_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 54000.0, "field": "sub_recorded_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 22500.0, "field": "difference", "tol_abs": 0.02, "type": "number"}, {"expect": ["ICINV-03"], "field": "difference_cause", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-026-intercompany-tieout-feb", "task_name": "cross_system/intercompany-tieout-feb", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"difference": 22500.0, "difference_cause": "ICINV-03 ($22,500) — received by the subsidiary 2026-03-01 but not yet booked (in transit)", "parent_ar_balance": 76500.0, "sub_recorded_balance": 54000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-027-intercompany-tieout-feb-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of cross_system/intercompany-tieout-feb via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#7; chaos primitive: system A learned the truth before system B", "provenance": "variant", "version": "1.0.0", "walk_len": 5, "walk_servers": ["email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 17:02**\n\nIntercompany with CES Direct for February. What do we say they owe, what do they say, and what's behind any difference?\n", "rubric": {"checks": {"answer_checks": [{"expect": 76500.0, "field": "parent_ar_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 54000.0, "field": "sub_recorded_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 22500.0, "field": "difference", "tol_abs": 0.02, "type": "number"}, {"expect": ["ICINV-03"], "field": "difference_cause", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-027-intercompany-tieout-feb-esc-burie-quiet", "task_name": "cross_system/intercompany-tieout-feb-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"combined_exposure": 113799.19, "erp_open_balance": 94749.19, "subsidiary_net_balance": 19050.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-028-total-ar-adventure-group", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "novel: data-chaos pattern 'subsidiary AR lives outside the ERP' (research/domain-workflows.md §3); FB has no cross-system tasks", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["books", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 10:41**\n\nAdventure Works — what's our total exposure to them across the group? Treasury wants one number they can defend.\n", "rubric": {"checks": {"answer_checks": [{"expect": 94749.19, "field": "erp_open_balance", "gt_sql": "SELECT ROUND(SUM(amount-settled),2) FROM erp_cust_trans WHERE account='SYNCUS-0002' AND txn_type='Invoice' AND closed=0", "tol_abs": 0.02, "type": "number"}, {"expect": 19050.0, "field": "subsidiary_net_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 113799.19, "field": "combined_exposure", "tol_abs": 0.04, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "books"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-028-total-ar-adventure-group", "task_name": "cross_system/total-ar-adventure-group", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"combined_exposure": 119999.19, "erp_open_balance": 94749.19, "offbook_invoice_amount": 6200.0, "subsidiary_net_balance": 19050.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "harder", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-029-total-ar-adventure-group-v2", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "grow: escalated from cross_system/total-ar-adventure-group (haiku passed 2/2 -> too_easy). Escalations: +sheet-only side invoice (email-corroborated), +stale AR-summary distractor that must be ignored, walk 6->10.", "provenance": "ported", "version": "1.0.0", "walk_len": 10, "walk_servers": ["books", "email", "erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 10:41**\n\nAdventure Works — total group exposure, one defensible number. Dana mentioned paperwork floating around from the CES cutover, so make sure nothing's missed or double-counted, and don't trust every file on the drive.\n", "rubric": {"checks": {"answer_checks": [{"expect": 94749.19, "field": "erp_open_balance", "gt_sql": "SELECT ROUND(SUM(amount-settled),2) FROM erp_cust_trans WHERE account='SYNCUS-0002' AND txn_type='Invoice' AND closed=0", "tol_abs": 0.02, "type": "number"}, {"expect": 19050.0, "field": "subsidiary_net_balance", "tol_abs": 0.02, "type": "number"}, {"expect": 6200.0, "field": "offbook_invoice_amount", "tol_abs": 0.01, "type": "number"}, {"expect": 119999.19, "field": "combined_exposure", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "books", "sheets"], "type": "required_servers"}, {"n": 2, "server": "books", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-029-total-ar-adventure-group-v2", "task_name": "cross_system/total-ar-adventure-group-v2", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"erp_live_total": 127450.0, "stale_customers": "Tailwind Bikes, Ostara Labs", "tracker_cached_total": 118400.0, "tracker_rows_sum": 228138.01}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "cross_system", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-030-tracker-formula-drift", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "chaos patterns 'spreadsheet version drift' + '1 in 3 CFOs distrust close data' (research/domain-workflows.md §3); enabled by sheets formula-cell support (Graph serves cached values, like real Excel)", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["erp", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 08:31**\n\nThe CFO quoted a watchlist number this morning and Casey says it feels low. Can you check it before I reply to him? Tell me what the number should be.\n", "rubric": {"checks": {"answer_checks": [{"expect": 118400.0, "field": "tracker_cached_total", "tol_abs": 0.02, "type": "number"}, {"expect": 228138.01, "field": "tracker_rows_sum", "tol_abs": 0.02, "type": "number"}, {"expect": 127450.0, "field": "erp_live_total", "tol_abs": 0.02, "type": "number"}, {"expect": ["tailwind", "ostara"], "field": "stale_customers", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "sheets"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-030-tracker-formula-drift", "task_name": "cross_system/tracker-formula-drift", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"as_of_date": "2026-03-02", "total_overdue_ap": 30616849.18}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-031-ap-overdue-usmf", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:erp_qa/AP Overdue (user-brief example; GT computed in-world)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 08:52**\n\nCash call at 11. How much of our AP in USMF is actually overdue as of today? I need the number, not a range.\n", "rubric": {"checks": {"answer_checks": [{"expect": 30616849.18, "field": "total_overdue_ap", "gt_sql": "SELECT ROUND(SUM(amount-settled),2) FROM erp_vend_trans WHERE txn_type='Invoice' AND closed=0 AND due_date<'2026-03-02'", "tol_abs": 0.05, "type": "number"}, {"expect": ["2026-03-02"], "field": "as_of_date", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-031-ap-overdue-usmf", "task_name": "erp_qa/ap-overdue-usmf", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"customer_account": "SYNCUS-0069", "open_invoice_count": 4, "outstanding_balance": 121321.26}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-032-ar-balance-fourthcoffee-east", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:erp_qa/Customer Balance (user-brief example rephrased onto in-world entity)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 09:14**\n\nFourth Coffee East just called about their account. What are we carrying on them right now?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["SYNCUS-0069"], "field": "customer_account", "type": "contains_all"}, {"expect": 121321.26, "field": "outstanding_balance", "gt_sql": "SELECT ROUND(SUM(amount-settled),2) FROM erp_cust_trans WHERE account='SYNCUS-0069' AND txn_type='Invoice' AND closed=0", "tol_abs": 0.02, "type": "number"}, {"expect": 4, "field": "open_invoice_count", "gt_sql": "SELECT COUNT(*) FROM erp_cust_trans WHERE account='SYNCUS-0069' AND txn_type='Invoice' AND closed=0", "tol_abs": 0, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-032-ar-balance-fourthcoffee-east", "task_name": "erp_qa/ar-balance-fourthcoffee-east", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"customer_account": "SYNCUS-0069", "open_invoice_count": 4, "outstanding_balance": 121321.26}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-033-ar-balance-fourthcoffee-east-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of erp_qa/ar-balance-fourthcoffee-east via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark:erp_qa/Customer Balance (user-brief example rephrased onto in-world entity)", "provenance": "variant", "version": "1.0.0", "walk_len": 4, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 09:14**\n\nFourth Coffee East just called about their account. What are we carrying on them right now?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["SYNCUS-0069"], "field": "customer_account", "type": "contains_all"}, {"expect": 121321.26, "field": "outstanding_balance", "gt_sql": "SELECT ROUND(SUM(amount-settled),2) FROM erp_cust_trans WHERE account='SYNCUS-0069' AND txn_type='Invoice' AND closed=0", "tol_abs": 0.02, "type": "number"}, {"expect": 4, "field": "open_invoice_count", "gt_sql": "SELECT COUNT(*) FROM erp_cust_trans WHERE account='SYNCUS-0069' AND txn_type='Invoice' AND closed=0", "tol_abs": 0, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-033-ar-balance-fourthcoffee-east-esc-burie-quiet", "task_name": "erp_qa/ar-balance-fourthcoffee-east-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-034-cash-disc-fourthcoffee-east/documents/policy--cash-discount-capture.md", "task_files/lgr100-034-cash-disc-fourthcoffee-east/inputs/statement-fourthcoffee-east-2026-02.csv"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"has_active_discount": "yes", "potential_discount_usd": 437.11, "qualifying_invoices": "VINV-900001, VINV-900002"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-034-cash-disc-fourthcoffee-east", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:erp_qa/Cash Discount + domain-workflows discount-capture runbook", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 08:20**\n\nFourth Coffee East's AR desk says we're leaving early-payment money on the table. Are we? If so, what can we still capture if I pay today?\n", "rubric": {"checks": {"answer_checks": [{"expect": "yes", "field": "has_active_discount", "type": "yes_no"}, {"expect": ["VINV-900001", "VINV-900002"], "field": "qualifying_invoices", "type": "contains_all"}, {"expect": 437.11, "field": "potential_discount_usd", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-034-cash-disc-fourthcoffee-east", "task_name": "erp_qa/cash-disc-fourthcoffee-east", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-035-cash-disc-fourthcoffee-east-esc-burie-quiet/documents/policy--cash-discount-capture.md", "task_files/lgr100-035-cash-disc-fourthcoffee-east-esc-burie-quiet/inputs/statement-fourthcoffee-east-2026-02.csv"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"has_active_discount": "yes", "potential_discount_usd": 437.11, "qualifying_invoices": "VINV-900001, VINV-900002"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-035-cash-disc-fourthcoffee-east-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of erp_qa/cash-disc-fourthcoffee-east via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark:erp_qa/Cash Discount + domain-workflows discount-capture runbook", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 08:20**\n\nFourth Coffee East's AR desk says we're leaving early-payment money on the table. Are we? If so, what can we still capture if I pay today?\n", "rubric": {"checks": {"answer_checks": [{"expect": "yes", "field": "has_active_discount", "type": "yes_no"}, {"expect": ["VINV-900001", "VINV-900002"], "field": "qualifying_invoices", "type": "contains_all"}, {"expect": 437.11, "field": "potential_discount_usd", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-035-cash-disc-fourthcoffee-east-esc-burie-quiet", "task_name": "erp_qa/cash-disc-fourthcoffee-east-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-036-collections-sparrow/documents/policy--collections-dunning-runbook.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"collection_letter_level": "2", "last_payment_amount": 22400.0, "last_payment_date": "2026-02-20", "open_balance": 28450.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-036-collections-sparrow", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:erp_qa/Collection Letters + Payment History; customer injected by task seed (Contoso demo US-008)", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 16:40**\n\nSparrow Retail is on tomorrow's agenda. Where are we with them, and have they paid us anything lately?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["2"], "field": "collection_letter_level", "type": "contains_all"}, {"expect": ["2026-02-20"], "field": "last_payment_date", "type": "contains_all"}, {"expect": 22400.0, "field": "last_payment_amount", "tol_abs": 0.01, "type": "number"}, {"expect": 28450.0, "field": "open_balance", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-036-collections-sparrow", "task_name": "erp_qa/collections-sparrow", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-037-collections-sparrow-esc-burie-quiet/documents/policy--collections-dunning-runbook.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"collection_letter_level": "2", "last_payment_amount": 22400.0, "last_payment_date": "2026-02-20", "open_balance": 28450.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-037-collections-sparrow-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of erp_qa/collections-sparrow via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark:erp_qa/Collection Letters + Payment History; customer injected by task seed (Contoso demo US-008)", "provenance": "variant", "version": "1.0.0", "walk_len": 5, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 16:40**\n\nSparrow Retail is on tomorrow's agenda. Where are we with them, and have they paid us anything lately?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["2"], "field": "collection_letter_level", "type": "contains_all"}, {"expect": ["2026-02-20"], "field": "last_payment_date", "type": "contains_all"}, {"expect": 22400.0, "field": "last_payment_amount", "tol_abs": 0.01, "type": "number"}, {"expect": 28450.0, "field": "open_balance", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-037-collections-sparrow-esc-burie-quiet", "task_name": "erp_qa/collections-sparrow-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"credit_limit": 25000, "customer_name": "A. Datum Corporation"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "easy", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-038-credit-limit-adatum", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:erp_qa/Credit Limit (query verbatim; GT recomputed in-world)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 10:02**\n\nSales wants to push an order through for SYNCUS-0001. What's their limit?\n", "rubric": {"checks": {"answer_checks": [{"expect": 25000, "field": "credit_limit", "tol_abs": 0.01, "type": "number"}, {"expect": ["datum"], "field": "customer_name", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-038-credit-limit-adatum", "task_name": "erp_qa/credit-limit-adatum", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"credit_limit": 25000, "customer_name": "A. Datum Corporation"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "easy", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-039-credit-limit-adatum-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of erp_qa/credit-limit-adatum via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark:erp_qa/Credit Limit (query verbatim; GT recomputed in-world)", "provenance": "variant", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 10:02**\n\nSales wants to push an order through for SYNCUS-0001. What's their limit?\n", "rubric": {"checks": {"answer_checks": [{"expect": 25000, "field": "credit_limit", "tol_abs": 0.01, "type": "number"}, {"expect": ["datum"], "field": "customer_name", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-039-credit-limit-adatum-esc-burie-quiet", "task_name": "erp_qa/credit-limit-adatum-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"invoice_count": 0, "invoice_numbers": "none"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-040-due-next-week-adventure", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark pattern: deliberately-empty ground truths (hallucination traps)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 09:30**\n\nBuilding next week's call list. Anything of Adventure Works Cycles' (SYNCUS-0002) falling due in the next seven days?\n", "rubric": {"checks": {"answer_checks": [{"expect": 0, "field": "invoice_count", "gt_sql": "SELECT COUNT(*) FROM erp_cust_trans WHERE account='SYNCUS-0002' AND txn_type='Invoice' AND closed=0 AND due_date>'2026-03-02' AND due_date<='2026-03-09'", "tol_abs": 0, "type": "number"}, {"field": "invoice_numbers", "type": "none_answer"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-040-due-next-week-adventure", "task_name": "erp_qa/due-next-week-adventure", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"invoice_count": 0, "invoice_numbers": "none"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-041-due-next-week-adventure-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of erp_qa/due-next-week-adventure via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark pattern: deliberately-empty ground truths (hallucination traps)", "provenance": "variant", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams 09:30**\n\nBuilding next week's call list. Anything of Adventure Works Cycles' (SYNCUS-0002) falling due in the next seven days?\n", "rubric": {"checks": {"answer_checks": [{"expect": 0, "field": "invoice_count", "gt_sql": "SELECT COUNT(*) FROM erp_cust_trans WHERE account='SYNCUS-0002' AND txn_type='Invoice' AND closed=0 AND due_date>'2026-03-02' AND due_date<='2026-03-09'", "tol_abs": 0, "type": "number"}, {"field": "invoice_numbers", "type": "none_answer"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-041-due-next-week-adventure-esc-burie-quiet", "task_name": "erp_qa/due-next-week-adventure-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"current_not_due": 5817896.81, "days_1_30": 3917195.87, "days_31_60": 2314492.37, "days_61_90": 2499959.29, "days_over_90": 16369769.19}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-042-aged-balance-12", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark erp_qa (AR / Aged Balance); question verbatim, ground truth recomputed in-world (docs/AUDIT.md A3)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams**\n\nWhat is the total balance of unpaid transactions for all customers by aging bucket within USMF as of March 2, 2026?\n", "rubric": {"checks": {"answer_checks": [{"expect": 5817896.81, "field": "current_not_due", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) <= 0", "tol_abs": 0.02, "type": "number"}, {"expect": 3917195.87, "field": "days_1_30", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 0 AND julianday('2026-03-02') - julianday(due_date) <= 30", "tol_abs": 0.02, "type": "number"}, {"expect": 2314492.37, "field": "days_31_60", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 30 AND julianday('2026-03-02') - julianday(due_date) <= 60", "tol_abs": 0.02, "type": "number"}, {"expect": 2499959.29, "field": "days_61_90", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 60 AND julianday('2026-03-02') - julianday(due_date) <= 90", "tol_abs": 0.02, "type": "number"}, {"expect": 16369769.19, "field": "days_over_90", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 90", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-042-aged-balance-12", "task_name": "erp_qa_fb/aged-balance-12", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"not_yet_due": 376.41, "over_90_days": 0.0, "total_past_due": 1084.11}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-043-aged-balance-6", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark erp_qa (AR / Aged Balance); question verbatim, ground truth recomputed in-world (docs/AUDIT.md A3)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams**\n\nCould you provide the breakdown of overdue receivables for Birch Company in USMF as of March 2, 2026?\n", "rubric": {"checks": {"answer_checks": [{"expect": 1084.11, "field": "total_past_due", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE account IN ('US-027') AND txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 0", "tol_abs": 0.02, "type": "number"}, {"expect": 376.41, "field": "not_yet_due", "tol_abs": 0.02, "type": "number"}, {"expect": 0.0, "field": "over_90_days", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-043-aged-balance-6", "task_name": "erp_qa_fb/aged-balance-6", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"invoice_count": 434, "past_due_180_plus_total": 9903381.01}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-044-aged-balance-7", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark erp_qa (AR / Aged Balance); question verbatim, ground truth recomputed in-world (docs/AUDIT.md A3)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams**\n\nWhat is the total balance of unpaid customer receivables transactions for 180+ days in USMF as of March 2, 2026\n", "rubric": {"checks": {"answer_checks": [{"expect": 434, "field": "invoice_count", "gt_sql": "SELECT COUNT(*) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 180", "tol_abs": 0, "type": "number"}, {"expect": 9903381.01, "field": "past_due_180_plus_total", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 180", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-044-aged-balance-7", "task_name": "erp_qa_fb/aged-balance-7", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"not_yet_due": 376.41, "over_90_days": 0.0, "total_past_due": 1084.11}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-045-aged-balance-8", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark erp_qa (AR / Aged Balance); question verbatim, ground truth recomputed in-world (docs/AUDIT.md A3)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Casey Morgan · AR & Collections · Teams**\n\nWhat is the aged balance for Birch Company in USMF as of March 2, 2026?\n", "rubric": {"checks": {"answer_checks": [{"expect": 1084.11, "field": "total_past_due", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_cust_trans WHERE account IN ('US-027') AND txn_type='Invoice' AND closed=0 AND julianday('2026-03-02') - julianday(due_date) > 0", "tol_abs": 0.02, "type": "number"}, {"expect": 376.41, "field": "not_yet_due", "tol_abs": 0.02, "type": "number"}, {"expect": 0.0, "field": "over_90_days", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-045-aged-balance-8", "task_name": "erp_qa_fb/aged-balance-8", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"invoices_due_this_week_count": 58, "invoices_due_this_week_total": 1540586.74}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erp_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-046-ap-invoices-3", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark erp_qa (AP / AP Invoices); question verbatim, ground truth recomputed in-world (docs/AUDIT.md A3)", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams**\n\nWhich vendor invoices are due for payment this week in USMF?\n", "rubric": {"checks": {"answer_checks": [{"expect": 58, "field": "invoices_due_this_week_count", "gt_sql": "SELECT COUNT(*) FROM erp_vend_trans WHERE txn_type='Invoice' AND closed=0 AND due_date BETWEEN '2026-03-02' AND '2026-03-08'", "tol_abs": 0, "type": "number"}, {"expect": 1540586.74, "field": "invoices_due_this_week_total", "gt_sql": "SELECT ROUND(COALESCE(SUM(amount-settled),0),2) FROM erp_vend_trans WHERE txn_type='Invoice' AND closed=0 AND due_date BETWEEN '2026-03-02' AND '2026-03-08'", "tol_abs": 0.02, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-046-ap-invoices-3", "task_name": "erp_qa_fb/ap-invoices-3", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 638, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1595, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 282, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-2771.11)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 664198.04, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 6862.0, "orders_accepted": 29, "orders_rejected": 0, "units_manufactured": 282, "units_purchased": 1595}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-047-2141-hard-15-parallel-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2141_hard_15_parallel_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 94, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nEnsure all these Portable Power Station 3kWh customer orders are supplied on schedule:\n\n- Solace Collective: 19 units due in 8 days (pretax budget cap $55,306)\n- Stonewall Research: 22 units due in 8 days (pretax budget cap $67,564)\n- Canton Trust: 22 units due in 8 days (pretax budget cap $65,613)\n- Clearwater Productions: 18 units due in 8 days (pretax budget cap $51,374)\n- Cobalt Innovations: 19 units due in 8 days (pretax budget cap $56,217)\n- Polar Cooperative: 20 units due in 8 days (pretax budget cap $61,053)\n- Marble Interactive: 32 units due in 9 days (pretax budget cap $95,218)\n- Sapphire Boutique: 19 units due in 9 days (pretax budget cap $56,037)\n- Lance Consortium: 18 units due in 9 days (pretax budget cap $51,002)\n- Noble Initiative: 18 units due in 9 days (pretax budget cap $50,942)\n- Crown Pictures: 18 units due in 9 days (pretax budget cap $53,056)\n- Lakewood Refinery: 19 units due in 9 days (pretax budget cap $58,253)\n- Haven Studios North: 25 units due in 9 days (pretax budget cap $76,512)\n- Bridgeway Chambers: 20 units due in 10 days (pretax budget cap $61,820)\n- Cardinal Dynamics: 32 units due in 10 days (pretax budget cap $98,260)\n- Oakmont Foundry: 21 units due in 11 days (pretax budget cap $59,760)\n- Catalyst Bureau: 25 units due in 11 days (pretax budget cap $71,903)\n- Alpine Theater: 18 units due in 11 days (pretax budget cap $53,484)\n- Lattice Technologies: 19 units due in 11 days (pretax budget cap $53,913)\n- Horizon Publishing: 22 units due in 11 days (pretax budget cap $66,084)\n- Globe Conservatory: 19 units due in 11 days (pretax budget cap $58,412)\n- Ironwood Forum: 31 units due in 12 days (pretax budget cap $96,027)\n- Brookfield Robotics: 23 units due in 12 days (pretax budget cap $68,639)\n- Ridgeline Sciences: 19 units due in 12 days (pretax budget cap $58,583)\n- Blaze Alliance: 25 units due in 13 days (pretax budget cap $73,118)\n- Meridian Atelier: 24 units due in 13 days (pretax budget cap $69,836)\n- Compass Institute: 19 units due in 13 days (pretax budget cap $58,252)\n- Steel Outfitters: 32 units due in 13 days (pretax budget cap $95,959)\n- Trellis Workshop: 20 units due in 13 days (pretax budget cap $59,420)\n\nCurrent finished-goods stock is 256 units. If stock runs short, you can close the gap with finished-goods purchasing, in-house manufacturing, or a combination of the two.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 29.2% at selling price.\n\n## Background & Policy\n\n* Fulfill all customer orders while preserving as much shared workcenter capacity as possible for other scheduled work. If more than one feasible plan uses the same amount of workcenter capacity, keep new purchasing and manufacturing spend as low as possible.\n* Any units you cover through new buying or manufacturing count toward one combined portfolio that must still meet a minimum 29.2% new-spend margin at selling price.\n* Available finished stock can be used where it helps keep shared workcenter capacity open.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Before releasing anything, read the Internal Notes/comments on stock, customers, vendors, and workcenters.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 6862.0, "field": "assembly_cost", "tol_abs": 34.31, "type": "number"}, {"expect": 29, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 1595, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 282, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 638, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1595, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 282, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-2771.11)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 664198.04, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-047-2141-hard-15-parallel-subassemblies-branch-assigned", "task_name": "erpbench/2141-hard-15-parallel-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 480, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 3202, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 753, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-5368.26)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 225259.64, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 18323.0, "orders_accepted": 24, "orders_rejected": 0, "units_manufactured": 753, "units_purchased": 3202}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-048-2143-hard-15-parallel-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2143_hard_15_parallel_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 98, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nAll orders below for LFP Energy Storage Module 51.2V need confirmed supply coverage and scheduling.\n\n- Ashford Ventures: 18 units due in 8 days (pretax budget cap $105,479)\n- Spark Architects: 18 units due in 8 days (pretax budget cap $106,836)\n- Cipher Partners: 20 units due in 8 days (pretax budget cap $116,691)\n- Spectra Agency: 19 units due in 8 days (pretax budget cap $104,471)\n- Nexus Pictures: 21 units due in 8 days (pretax budget cap $117,741)\n- Arbor Group: 19 units due in 9 days (pretax budget cap $107,214)\n- Monarch Designs: 18 units due in 9 days (pretax budget cap $100,637)\n- Meridian Exchange: 21 units due in 9 days (pretax budget cap $123,501)\n- Summit Advisory: 18 units due in 9 days (pretax budget cap $106,895)\n- Raven Workspaces: 19 units due in 10 days (pretax budget cap $113,369)\n- Metro Creative: 18 units due in 10 days (pretax budget cap $105,245)\n- Flux Refinery: 20 units due in 10 days (pretax budget cap $118,815)\n- Zenith Analytics: 19 units due in 10 days (pretax budget cap $106,533)\n- Axis Collective East: 18 units due in 10 days (pretax budget cap $106,262)\n- Sterling Solutions Group: 22 units due in 10 days (pretax budget cap $128,995)\n- Haven Workshop: 32 units due in 11 days (pretax budget cap $177,998)\n- Alpine Hub: 18 units due in 11 days (pretax budget cap $101,263)\n- Anvil Clinics: 19 units due in 11 days (pretax budget cap $110,223)\n- Dune Productions: 18 units due in 11 days (pretax budget cap $101,580)\n- Echo Institute: 21 units due in 12 days (pretax budget cap $121,275)\n- Orbital Society: 25 units due in 12 days (pretax budget cap $141,525)\n- Sapphire Enterprises: 18 units due in 12 days (pretax budget cap $105,679)\n- Matrix Media: 18 units due in 13 days (pretax budget cap $104,802)\n- Stratos Academy: 23 units due in 13 days (pretax budget cap $135,124)\n\nCurrent finished-goods stock is 192 units. Any shortfall can be handled with finished-goods buying, in-house manufacturing, or a mix that still satisfies policy.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 27.1% at selling price.\n\n## Background & Policy\n\n* Cover every customer order while using as little shared workcenter capacity as practical. If multiple feasible plans use the same amount of workcenter capacity, keep new purchasing and manufacturing spend as low as possible.\n* Any units you cover through new buying or manufacturing count toward one combined portfolio that must still meet a minimum 27.1% new-spend margin at selling price.\n* Use available finished stock where it helps preserve shared workcenter capacity.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Review Internal Notes/comments on stock, customers, vendors, and workcenters before you confirm entries in the ERP.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 18323.0, "field": "assembly_cost", "tol_abs": 91.615, "type": "number"}, {"expect": 24, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 3202, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 753, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 480, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 3202, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 753, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-5368.26)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 225259.64, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-048-2143-hard-15-parallel-subassemblies-branch-assigned", "task_name": "erpbench/2143-hard-15-parallel-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 520, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 2841, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 408, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-2860.79)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 431224.32, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 9928.0, "orders_accepted": 26, "orders_rejected": 0, "units_manufactured": 408, "units_purchased": 2841}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-049-2144-hard-15-parallel-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2144_hard_15_parallel_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 92, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nAll orders below for Portable Power Station 3kWh need confirmed supply coverage and scheduling.\n\n- Marble Observatory: 19 units due in 8 days (pretax budget cap $58,130)\n- Forge Dynamics: 18 units due in 8 days (pretax budget cap $57,406)\n- Aurora Collective East: 18 units due in 8 days (pretax budget cap $54,431)\n- Ember Arena: 18 units due in 8 days (pretax budget cap $55,690)\n- Crown Works: 21 units due in 8 days (pretax budget cap $65,303)\n- Hartland Theater: 18 units due in 8 days (pretax budget cap $55,782)\n- Haven Productions: 19 units due in 8 days (pretax budget cap $57,965)\n- Monarch Reserve: 20 units due in 8 days (pretax budget cap $60,246)\n- Metro Atelier: 24 units due in 9 days (pretax budget cap $76,407)\n- Meridian Publishing: 18 units due in 9 days (pretax budget cap $52,527)\n- Summit Advisory: 21 units due in 9 days (pretax budget cap $63,638)\n- Oxide Studios North: 18 units due in 9 days (pretax budget cap $55,107)\n- Oakmont Workshop: 23 units due in 10 days (pretax budget cap $67,454)\n- Flint Solutions Group: 20 units due in 10 days (pretax budget cap $64,080)\n- Helix Studios West: 27 units due in 10 days (pretax budget cap $79,359)\n- Sterling Pavilion: 19 units due in 11 days (pretax budget cap $56,080)\n- Quantum Architects: 18 units due in 11 days (pretax budget cap $54,496)\n- Cardinal Trading Co: 18 units due in 11 days (pretax budget cap $53,629)\n- Aegis Society: 22 units due in 11 days (pretax budget cap $68,840)\n- Ironside Alliance: 20 units due in 12 days (pretax budget cap $63,101)\n- Northbridge Manufactory: 21 units due in 12 days (pretax budget cap $64,478)\n- Crest Sciences: 18 units due in 13 days (pretax budget cap $54,838)\n- Baltic Consortium: 23 units due in 13 days (pretax budget cap $67,443)\n- Trident Technologies: 21 units due in 13 days (pretax budget cap $62,643)\n- Axis Partners: 19 units due in 13 days (pretax budget cap $57,175)\n- Scion Museum: 19 units due in 13 days (pretax budget cap $56,782)\n\nOn-hand finished stock covers 208 units. If stock runs short, you can close the gap with finished-goods purchasing, in-house manufacturing, or a combination of the two.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 26.9% at selling price.\n\n## Background & Policy\n\n* Fulfill all customer orders while preserving as much shared workcenter capacity as possible for other scheduled work. If more than one feasible plan uses the same amount of workcenter capacity, keep new purchasing and manufacturing spend as low as possible.\n* The combined units covered through new purchasing or manufacturing must clear at least 26.9% portfolio-level new-spend margin at selling price.\n* Use available finished stock where it helps preserve shared workcenter capacity.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Check Internal Notes/comments on stock, customers, vendors, and workcenters before you release anything.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 9928.0, "field": "assembly_cost", "tol_abs": 49.64, "type": "number"}, {"expect": 26, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 2841, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 408, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 520, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 2841, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 408, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-2860.79)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 431224.32, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-049-2144-hard-15-parallel-subassemblies-branch-assigned", "task_name": "erpbench/2144-hard-15-parallel-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 690, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 4752, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 597, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-16199.21)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 2566256.86, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 14527.0, "orders_accepted": 30, "orders_rejected": 0, "units_manufactured": 597, "units_purchased": 4752}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-050-2146-hard-15-parallel-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2146_hard_15_parallel_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 104, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nAll orders below for Modular Rack Battery System 20kWh need confirmed supply coverage and scheduling.\n\n- Beacon Arena: 20 units due in 8 days (pretax budget cap $352,056)\n- Dune Enterprises: 19 units due in 8 days (pretax budget cap $327,706)\n- Meridian Advisory: 18 units due in 8 days (pretax budget cap $303,628)\n- Aether Consortium: 20 units due in 8 days (pretax budget cap $359,031)\n- Sierra Clinics: 19 units due in 9 days (pretax budget cap $342,730)\n- Axis Robotics: 20 units due in 9 days (pretax budget cap $359,410)\n- Trident Research: 29 units due in 9 days (pretax budget cap $505,931)\n- Westfield Archive: 18 units due in 9 days (pretax budget cap $314,483)\n- Bridgeway Trading Co: 22 units due in 9 days (pretax budget cap $364,608)\n- Quantum Initiative: 19 units due in 10 days (pretax budget cap $322,719)\n- Catalyst Council: 20 units due in 10 days (pretax budget cap $352,781)\n- Cipher Lyceum: 23 units due in 10 days (pretax budget cap $389,731)\n- Blaze Greenhouse: 19 units due in 10 days (pretax budget cap $328,371)\n- Peak Workspaces: 22 units due in 10 days (pretax budget cap $363,871)\n- Ivory Studios North: 25 units due in 10 days (pretax budget cap $423,873)\n- Hartland Theater: 18 units due in 11 days (pretax budget cap $311,063)\n- Oakmont Society: 28 units due in 11 days (pretax budget cap $482,909)\n- Citadel Studios: 32 units due in 11 days (pretax budget cap $556,306)\n- Grove Bureau: 24 units due in 11 days (pretax budget cap $412,467)\n- Gateway Institute: 31 units due in 11 days (pretax budget cap $519,842)\n- Flux Chambers: 21 units due in 11 days (pretax budget cap $356,602)\n- Arrow Agency: 21 units due in 12 days (pretax budget cap $356,370)\n- Nexus Group: 25 units due in 12 days (pretax budget cap $416,876)\n- Spark Museum: 32 units due in 12 days (pretax budget cap $543,868)\n- Raven Studios West: 25 units due in 12 days (pretax budget cap $424,692)\n- Quartz Pictures: 25 units due in 13 days (pretax budget cap $429,558)\n- Metro Studios South: 20 units due in 13 days (pretax budget cap $334,544)\n- Ridge Workshop: 18 units due in 13 days (pretax budget cap $300,811)\n- Aegis Conservatory: 25 units due in 13 days (pretax budget cap $430,898)\n- Cobalt Pavilion: 32 units due in 13 days (pretax budget cap $529,235)\n\nWe have 276 finished units on hand. If stock is not enough, you can cover the gap by buying finished goods, building in-house, or using a mix of both.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 25% at selling price.\n\n## Background & Policy\n\n* Cover every customer order while using as little shared workcenter capacity as practical. If multiple feasible plans use the same amount of workcenter capacity, keep new purchasing and manufacturing spend as low as possible.\n* Any units you cover through new buying or manufacturing count toward one combined portfolio that must still meet a minimum 25% new-spend margin at selling price.\n* Use available finished stock where it helps preserve shared workcenter capacity.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Check Internal Notes/comments on stock, customers, vendors, and workcenters before you release anything.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 14527.0, "field": "assembly_cost", "tol_abs": 72.635, "type": "number"}, {"expect": 30, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 4752, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 597, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 690, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 4752, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 597, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-16199.21)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 2566256.86, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-050-2146-hard-15-parallel-subassemblies-branch-assigned", "task_name": "erpbench/2146-hard-15-parallel-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 775, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 5820, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 624, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-5081.48)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 1101367.17, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 15184.0, "orders_accepted": 31, "orders_rejected": 0, "units_manufactured": 624, "units_purchased": 5820}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-051-2147-hard-15-parallel-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2147_hard_15_parallel_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 99, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nPrepare a fulfillment strategy for all listed LFP Energy Storage Module 51.2V orders to meet due dates and avoid shortages.\n\n- Metro Greenhouse: 19 units due in 8 days (pretax budget cap $103,396)\n- Velocity Supply Co: 18 units due in 8 days (pretax budget cap $96,004)\n- Cipher Pictures: 18 units due in 8 days (pretax budget cap $94,282)\n- Mosaic Collective East: 27 units due in 9 days (pretax budget cap $152,485)\n- Flux Creative: 32 units due in 9 days (pretax budget cap $170,208)\n- Ember Brands: 28 units due in 9 days (pretax budget cap $153,141)\n- Grove Collective: 18 units due in 9 days (pretax budget cap $96,124)\n- Evergreen Lyceum: 30 units due in 9 days (pretax budget cap $165,761)\n- Ridgeline Solutions Group: 28 units due in 9 days (pretax budget cap $146,640)\n- Quartz Studios: 19 units due in 9 days (pretax budget cap $103,705)\n- Aurora Trading Co: 20 units due in 10 days (pretax budget cap $113,434)\n- Raven Studios East: 18 units due in 10 days (pretax budget cap $101,750)\n- Horizon Agency: 20 units due in 10 days (pretax budget cap $103,718)\n- Ridge Forum: 21 units due in 11 days (pretax budget cap $111,977)\n- Globe Reserve: 26 units due in 11 days (pretax budget cap $136,052)\n- Aether Interactive: 32 units due in 11 days (pretax budget cap $168,986)\n- Atlas Office: 28 units due in 11 days (pretax budget cap $151,150)\n- Matrix Media: 23 units due in 11 days (pretax budget cap $128,934)\n- Stonewall Dynamics: 18 units due in 11 days (pretax budget cap $94,525)\n- Citadel Group: 32 units due in 11 days (pretax budget cap $181,044)\n- Lumen Manufactory: 23 units due in 12 days (pretax budget cap $126,189)\n- Gateway Conservatory: 20 units due in 12 days (pretax budget cap $107,844)\n- Comet Fabricators: 25 units due in 12 days (pretax budget cap $139,441)\n- Steel Analytics: 32 units due in 12 days (pretax budget cap $178,581)\n- Beacon Outfitters: 20 units due in 13 days (pretax budget cap $104,708)\n- Indigo Holdings: 20 units due in 13 days (pretax budget cap $103,868)\n- Monarch Advisory: 32 units due in 13 days (pretax budget cap $178,229)\n- Pinnacle Ventures: 32 units due in 13 days (pretax budget cap $181,485)\n- Apex Pavilion: 32 units due in 13 days (pretax budget cap $170,611)\n- Oxide Museum: 32 units due in 13 days (pretax budget cap $175,796)\n- Catalyst Guild: 32 units due in 13 days (pretax budget cap $168,880)\n\nWe have 310 finished units on hand. Any shortfall can be handled with finished-goods buying, in-house manufacturing, or a mix that still satisfies policy.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 27.4% at selling price.\n\n## Background & Policy\n\n* Get every customer order covered while keeping as much shared workcenter capacity open as possible. Keep new purchasing and manufacturing spend as low as possible when workcenter-capacity use ties.\n* Any units you cover through new buying or manufacturing count toward one combined portfolio that must still meet a minimum 27.4% new-spend margin at selling price.\n* Use available finished stock where it helps preserve shared workcenter capacity.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Before releasing anything, read the Internal Notes/comments on stock, customers, vendors, and workcenters.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 15184.0, "field": "assembly_cost", "tol_abs": 75.92, "type": "number"}, {"expect": 31, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 5820, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 624, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 775, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 5820, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 624, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-5081.48)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 1101367.17, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-051-2147-hard-15-parallel-subassemblies-branch-assigned", "task_name": "erpbench/2147-hard-15-parallel-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 806, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 2755, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 645, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-4284.53)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 841468.02, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 15695.0, "orders_accepted": 31, "orders_rejected": 0, "units_manufactured": 645, "units_purchased": 2755}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-052-2149-hard-15-parallel-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2149_hard_15_parallel_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 98, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nThe following customer orders for Solar Storage Battery Wall-Mount require confirmed supply to proceed with fulfillment:\n\n- Globe Media: 18 units due in 8 days (pretax budget cap $84,262)\n- Baseline Bureau: 29 units due in 8 days (pretax budget cap $128,511)\n- Keystone Trading Co: 19 units due in 8 days (pretax budget cap $90,783)\n- Blaze Society: 27 units due in 9 days (pretax budget cap $128,077)\n- Bridgeway Reserve: 21 units due in 9 days (pretax budget cap $99,938)\n- Quantum Atelier: 25 units due in 9 days (pretax budget cap $111,546)\n- Terra Systems: 32 units due in 10 days (pretax budget cap $145,536)\n- Ivory Forum: 24 units due in 10 days (pretax budget cap $114,766)\n- Polar Foundry: 18 units due in 10 days (pretax budget cap $86,070)\n- Mosaic Robotics: 22 units due in 10 days (pretax budget cap $97,604)\n- Pinnacle Guild: 22 units due in 10 days (pretax budget cap $105,497)\n- Trellis Institute: 22 units due in 10 days (pretax budget cap $101,082)\n- Copper Technologies: 19 units due in 11 days (pretax budget cap $86,294)\n- Horizon Studios South: 29 units due in 11 days (pretax budget cap $130,871)\n- Slate Manufactory: 25 units due in 11 days (pretax budget cap $113,789)\n- Indigo Studios East: 20 units due in 11 days (pretax budget cap $94,017)\n- Stonewall Conservatory: 19 units due in 11 days (pretax budget cap $85,194)\n- Oakmont Museum: 18 units due in 11 days (pretax budget cap $78,678)\n- Westfield Innovations: 19 units due in 11 days (pretax budget cap $89,195)\n- Axis Group: 28 units due in 12 days (pretax budget cap $128,546)\n- Skyline Alliance: 32 units due in 12 days (pretax budget cap $143,165)\n- Vertex Architects: 32 units due in 12 days (pretax budget cap $144,488)\n- Marble Greenhouse: 30 units due in 12 days (pretax budget cap $139,936)\n- Lumen Archive: 32 units due in 12 days (pretax budget cap $152,015)\n- Cosmo Designs: 32 units due in 12 days (pretax budget cap $140,371)\n- Dune Hub: 32 units due in 12 days (pretax budget cap $144,744)\n- Pivot Interactive: 32 units due in 12 days (pretax budget cap $142,630)\n- Noble Dynamics: 32 units due in 13 days (pretax budget cap $153,425)\n- Sentinel Analytics: 32 units due in 13 days (pretax budget cap $140,958)\n- Gateway Exchange: 32 units due in 13 days (pretax budget cap $150,107)\n- Echo Advisory: 32 units due in 13 days (pretax budget cap $140,604)\n\nCurrent finished-goods stock is 323 units. If stock runs short, you can close the gap with finished-goods purchasing, in-house manufacturing, or a combination of the two.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 26.3% at selling price.\n\n## Background & Policy\n\n* Get every customer order covered while keeping as much shared workcenter capacity open as possible. Keep new purchasing and manufacturing spend as low as possible when workcenter-capacity use ties.\n* The combined units covered through new purchasing or manufacturing must clear at least 26.3% portfolio-level new-spend margin at selling price.\n* Use available finished stock where it helps preserve shared workcenter capacity.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Review Internal Notes/comments on stock, customers, vendors, and workcenters before you confirm entries in the ERP.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 15695.0, "field": "assembly_cost", "tol_abs": 78.47500000000001, "type": "number"}, {"expect": 31, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 2755, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 645, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 806, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 2755, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 645, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-4284.53)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 841468.02, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-052-2149-hard-15-parallel-subassemblies-branch-assigned", "task_name": "erpbench/2149-hard-15-parallel-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 713, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1148, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 558, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-3905.52)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 780504.33, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 13578.0, "orders_accepted": 31, "orders_rejected": 0, "units_manufactured": 558, "units_purchased": 1148}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-053-2160-hard-17-shared-component-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2160_hard_17_shared_component_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 91, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nEach customer order below requires a fulfillment-ready supply solution before its delivery due date:\n\n- Garnet Guild: 32 units due in 8 days (pretax budget cap $132,393)\n- Terra Workshop: 21 units due in 8 days (pretax budget cap $91,224)\n- Titan Studios East: 18 units due in 8 days (pretax budget cap $73,335)\n- Eclipse Arena: 22 units due in 8 days (pretax budget cap $93,124)\n- Vantage Trust: 18 units due in 8 days (pretax budget cap $78,142)\n- Sapphire Engineering: 19 units due in 8 days (pretax budget cap $78,343)\n- Sterling Brands: 28 units due in 9 days (pretax budget cap $111,635)\n- Opal Exchange: 20 units due in 9 days (pretax budget cap $86,024)\n- Mosaic Partners: 18 units due in 9 days (pretax budget cap $77,949)\n- Quartz Pavilion: 18 units due in 9 days (pretax budget cap $77,673)\n- Northbridge Sciences: 21 units due in 10 days (pretax budget cap $86,710)\n- Compass Initiative: 29 units due in 10 days (pretax budget cap $120,205)\n- Brookfield Supply Co: 18 units due in 10 days (pretax budget cap $72,756)\n- Onyx Advisory: 29 units due in 10 days (pretax budget cap $126,004)\n- Spectra Workspaces: 19 units due in 11 days (pretax budget cap $80,170)\n- Ashford Consortium: 21 units due in 11 days (pretax budget cap $84,629)\n- Meridian Agency: 19 units due in 11 days (pretax budget cap $80,357)\n- Helix Publishing: 21 units due in 11 days (pretax budget cap $89,157)\n- Iron Research: 21 units due in 12 days (pretax budget cap $86,451)\n- Blaze Collective East: 24 units due in 12 days (pretax budget cap $102,909)\n- Cedar Creative: 18 units due in 12 days (pretax budget cap $76,823)\n- Echo Interactive: 19 units due in 12 days (pretax budget cap $81,671)\n- Nexus Innovations: 19 units due in 12 days (pretax budget cap $79,663)\n- Borough Archive: 32 units due in 12 days (pretax budget cap $138,783)\n- Gateway Greenhouse: 26 units due in 12 days (pretax budget cap $103,838)\n- Nimbus Forum: 27 units due in 12 days (pretax budget cap $116,870)\n- Baseline Conservatory: 21 units due in 12 days (pretax budget cap $88,888)\n- Crest Theater: 22 units due in 13 days (pretax budget cap $96,020)\n- Peak Productions: 29 units due in 13 days (pretax budget cap $116,357)\n- Horizon Designs: 32 units due in 13 days (pretax budget cap $137,250)\n- Ridge Group: 32 units due in 13 days (pretax budget cap $136,365)\n\nCurrent finished-goods stock is 286 units. If stock runs short, you can close the gap with finished-goods purchasing, in-house manufacturing, or a combination of the two.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 28.8% at selling price.\n\n## Background & Policy\n\n* Fulfill all orders while minimizing new spending on procurement and manufacturing.\n* Any units you cover through new buying or manufacturing count toward one combined portfolio that must still meet a minimum 28.8% new-spend margin at selling price.\n* Finance considers the existing stock as sunk cost.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Check Internal Notes/comments on stock, customers, vendors, and workcenters before you release anything.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 13578.0, "field": "assembly_cost", "tol_abs": 67.89, "type": "number"}, {"expect": 31, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 1148, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 558, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 713, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1148, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 558, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-3905.52)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 780504.33, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-053-2160-hard-17-shared-component-subassemblies-branch-assigned", "task_name": "erpbench/2160-hard-17-shared-component-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 750, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1052, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 468, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-5565.51)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 1437841.98, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 11388.0, "orders_accepted": 30, "orders_rejected": 0, "units_manufactured": 468, "units_purchased": 1052}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-054-2161-hard-17-shared-component-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2161_hard_17_shared_component_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 91, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nThe following customer orders for Dry Pipe Sprinkler System require confirmed supply to proceed with fulfillment:\n\n- Blaze Group: 19 units due in 8 days (pretax budget cap $113,022)\n- Forge Pavilion: 22 units due in 8 days (pretax budget cap $128,379)\n- Spark Arena: 20 units due in 8 days (pretax budget cap $119,848)\n- Matrix Observatory: 30 units due in 8 days (pretax budget cap $174,138)\n- Compass Works: 18 units due in 9 days (pretax budget cap $108,299)\n- Garnet Designs: 19 units due in 9 days (pretax budget cap $113,968)\n- Stonewall Labs: 32 units due in 9 days (pretax budget cap $182,563)\n- Echo Practice: 28 units due in 9 days (pretax budget cap $170,062)\n- Zenith Collective: 24 units due in 9 days (pretax budget cap $140,222)\n- Bayshore Manufactory: 19 units due in 10 days (pretax budget cap $109,360)\n- Cosmo Greenhouse: 32 units due in 11 days (pretax budget cap $194,916)\n- Chrome Interactive: 32 units due in 11 days (pretax budget cap $186,583)\n- Brookfield Gallery: 22 units due in 11 days (pretax budget cap $131,615)\n- Cobalt Archive: 21 units due in 11 days (pretax budget cap $129,146)\n- Drift Academy: 23 units due in 11 days (pretax budget cap $134,155)\n- Marble Systems: 19 units due in 11 days (pretax budget cap $113,553)\n- Evergreen Brands: 32 units due in 11 days (pretax budget cap $195,541)\n- Mosaic Analytics: 32 units due in 11 days (pretax budget cap $185,532)\n- Anvil Trading Co: 18 units due in 11 days (pretax budget cap $105,626)\n- Ledger Ventures: 18 units due in 11 days (pretax budget cap $111,601)\n- Crest Workshop: 22 units due in 12 days (pretax budget cap $132,165)\n- Osprey Dynamics: 19 units due in 12 days (pretax budget cap $110,270)\n- Trident Hub: 18 units due in 12 days (pretax budget cap $109,093)\n- Lakewood Workspaces: 21 units due in 12 days (pretax budget cap $127,574)\n- Metro Fabricators: 30 units due in 12 days (pretax budget cap $172,392)\n- Haven Technologies: 32 units due in 13 days (pretax budget cap $191,841)\n- Titan Reserve: 32 units due in 13 days (pretax budget cap $196,937)\n- Aegis Conservatory: 32 units due in 13 days (pretax budget cap $186,232)\n- Quartz Exchange: 32 units due in 13 days (pretax budget cap $194,566)\n- Ridgeline Enterprises: 32 units due in 13 days (pretax budget cap $191,183)\n\nWe have 300 finished units on hand. If stock is not enough, you can cover the gap by buying finished goods, building in-house, or using a mix of both.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 28.4% at selling price.\n\n## Background & Policy\n\n* Fulfill all orders while minimizing new spending on procurement and manufacturing.\n* Any units you cover through new buying or manufacturing count toward one combined portfolio that must still meet a minimum 28.4% new-spend margin at selling price.\n* Accounting treats the existing stock as sunk cost.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Review Internal Notes/comments on stock, customers, vendors, and workcenters before you confirm entries in the ERP.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 11388.0, "field": "assembly_cost", "tol_abs": 56.94, "type": "number"}, {"expect": 30, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 1052, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 468, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 750, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1052, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 468, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-5565.51)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 1437841.98, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-054-2161-hard-17-shared-component-subassemblies-branch-assigned", "task_name": "erpbench/2161-hard-17-shared-component-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 621, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1232, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 546, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-8748.9)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 970240.38, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 13286.0, "orders_accepted": 27, "orders_rejected": 0, "units_manufactured": 546, "units_purchased": 1232}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-055-2168-hard-17-shared-component-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2168_hard_17_shared_component_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 92, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nThe following Foam Deluge System Package customer orders are firm and must be supplied on schedule:\n\n- Zenith Cooperative: 18 units due in 8 days (pretax budget cap $175,969)\n- Conduit Studios North: 18 units due in 8 days (pretax budget cap $165,380)\n- Bayshore Office: 18 units due in 8 days (pretax budget cap $172,571)\n- Skyline Group: 32 units due in 8 days (pretax budget cap $296,374)\n- Canton Conservatory: 23 units due in 8 days (pretax budget cap $206,026)\n- Flint Media: 18 units due in 8 days (pretax budget cap $162,925)\n- Orbital Pavilion: 32 units due in 9 days (pretax budget cap $300,552)\n- Grove Workspaces: 18 units due in 9 days (pretax budget cap $166,463)\n- Crestview Ventures: 22 units due in 10 days (pretax budget cap $196,328)\n- Ledger Bureau: 18 units due in 10 days (pretax budget cap $176,275)\n- Sterling Creative: 23 units due in 11 days (pretax budget cap $206,322)\n- Ironwood Technologies: 20 units due in 11 days (pretax budget cap $193,330)\n- Ridge Manufactory: 20 units due in 11 days (pretax budget cap $192,570)\n- Echo Collective East: 20 units due in 11 days (pretax budget cap $188,568)\n- Quantum Solutions Group: 25 units due in 11 days (pretax budget cap $236,923)\n- Globe Publishing: 28 units due in 11 days (pretax budget cap $269,837)\n- Oxide Architects: 22 units due in 12 days (pretax budget cap $205,473)\n- Ember Fabricators: 26 units due in 12 days (pretax budget cap $243,660)\n- Keystone Agency: 23 units due in 12 days (pretax budget cap $207,313)\n- Opal Guild: 21 units due in 12 days (pretax budget cap $201,598)\n- Matrix Theater: 20 units due in 12 days (pretax budget cap $189,311)\n- Cardinal Exchange: 23 units due in 12 days (pretax budget cap $218,801)\n- Summit Designs: 23 units due in 13 days (pretax budget cap $207,853)\n- Mosaic Lyceum: 21 units due in 13 days (pretax budget cap $189,708)\n- Clearwater Clinics: 25 units due in 13 days (pretax budget cap $234,250)\n- Aegis Analytics: 32 units due in 13 days (pretax budget cap $298,511)\n- Onyx Boutique: 32 units due in 13 days (pretax budget cap $311,419)\n\nOn-hand finished stock covers 249 units. If stock is not enough, you can cover the gap by buying finished goods, building in-house, or using a mix of both.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 29.8% at selling price.\n\n## Background & Policy\n\n* Fulfill all orders while minimizing new spending on procurement and manufacturing.\n* Across all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 29.8% at selling price.\n* Accounting treats the existing stock as sunk cost.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Review Internal Notes/comments on stock, customers, vendors, and workcenters before you confirm entries in the ERP.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 13286.0, "field": "assembly_cost", "tol_abs": 66.43, "type": "number"}, {"expect": 27, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 1232, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 546, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 621, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1232, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 546, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-8748.9)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 970240.38, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-055-2168-hard-17-shared-component-subassemblies-branch-assigned", "task_name": "erpbench/2168-hard-17-shared-component-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 713, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1009, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 636, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-4143.04)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 670485.44, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}], "submit_answer": {"assembly_cost": 15476.0, "orders_accepted": 31, "orders_rejected": 0, "units_manufactured": 636, "units_purchased": 1009}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "erpbench", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-056-2170-hard-17-shared-component-subassemblies-branch-assigned", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of agentic-labs/erp-bench task 2170_hard_17_shared_component_subassemblies_branch_assigned; instruction carried verbatim, scenario seeded from its own environment/scenario_data.json, ground truth derived from its solution/optimal_plan.json. The source boots a real Odoo in Docker and grades with odoolib; this runs on the world's Odoo-shaped surface and grades the same judgement as deterministic SQL state assertions (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 98, "walk_servers": ["harness", "odoo"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Supply & Procurement · Teams**\n\nPrepare a fulfillment strategy for all listed Fire Alarm Control Panel FACP orders to meet due dates and avoid shortages.\n\n- Crown Trust: 18 units due in 8 days (pretax budget cap $76,849)\n- Bayshore Dynamics: 19 units due in 8 days (pretax budget cap $88,079)\n- Summit Lyceum: 24 units due in 8 days (pretax budget cap $108,956)\n- Aether Pavilion: 24 units due in 8 days (pretax budget cap $104,863)\n- Equinox Society: 18 units due in 8 days (pretax budget cap $78,824)\n- Sapphire Boutique: 18 units due in 9 days (pretax budget cap $80,144)\n- Quartz Engineering: 23 units due in 9 days (pretax budget cap $100,283)\n- Trident Labs: 21 units due in 9 days (pretax budget cap $95,134)\n- Scion Supply Co: 18 units due in 9 days (pretax budget cap $77,786)\n- Nimbus Solutions Group: 25 units due in 10 days (pretax budget cap $109,235)\n- Crestview Forum: 24 units due in 10 days (pretax budget cap $104,160)\n- Arbor Academy: 25 units due in 10 days (pretax budget cap $110,200)\n- Chrome Analytics: 22 units due in 10 days (pretax budget cap $97,251)\n- Atlas Agency: 21 units due in 10 days (pretax budget cap $89,989)\n- Osprey Studios West: 30 units due in 10 days (pretax budget cap $134,870)\n- Cipher Brands: 19 units due in 10 days (pretax budget cap $84,792)\n- Grove Manufactory: 18 units due in 11 days (pretax budget cap $80,277)\n- Compass Collective East: 21 units due in 11 days (pretax budget cap $94,971)\n- Pacific Workshop: 18 units due in 12 days (pretax budget cap $79,169)\n- Flint Observatory: 22 units due in 12 days (pretax budget cap $96,872)\n- Comet Arena: 18 units due in 12 days (pretax budget cap $79,861)\n- Eclipse Interactive: 22 units due in 12 days (pretax budget cap $94,032)\n- Axis Bureau: 19 units due in 12 days (pretax budget cap $80,809)\n- Borough Bazaar: 18 units due in 13 days (pretax budget cap $76,082)\n- Gateway Technologies: 30 units due in 13 days (pretax budget cap $130,290)\n- Ridge Workspaces: 18 units due in 13 days (pretax budget cap $81,936)\n- Cobalt Research: 32 units due in 13 days (pretax budget cap $141,224)\n- Catalyst Consortium: 32 units due in 13 days (pretax budget cap $140,917)\n- Quantum Holdings: 32 units due in 13 days (pretax budget cap $143,212)\n- Aurora Chambers: 32 units due in 13 days (pretax budget cap $144,242)\n- Ridgeline Pictures: 32 units due in 13 days (pretax budget cap $136,291)\n\nCurrent finished-goods stock is 286 units. If stock runs short, you can close the gap with finished-goods purchasing, in-house manufacturing, or a combination of the two.\n\nAcross all units covered through new purchasing or manufacturing, keep portfolio-level new-spend margin at or above 28.4% at selling price.\n\n## Background & Policy\n\n* Fulfill all orders while minimizing new spending on procurement and manufacturing.\n* The combined units covered through new purchasing or manufacturing must clear at least 28.4% portfolio-level new-spend margin at selling price.\n* Existing stock is a sunk cost and should not be treated as new spend.\n* You must create and confirm the necessary sales orders, purchase orders, and/or manufacturing orders.\n* Customer budgets are pre-tax amounts.\n* Link Sales Orders to the related Manufacturing Orders and Purchase Orders for traceability.\n* For finished goods POs, put the SO reference(s) (e.g. S00030 or S00030, S00031) into the origin field ('Source' in the UI).\n* For component POs, put the MO reference(s) (e.g. WH/MO/00010 or WH/MO/00010, WH/MO/00011) into the origin field.\n* For finished goods MOs, put the Sales Order reference (e.g. S00030) into the origin field ('Source' in the UI).\n* For subassembly or intermediate MOs, put the immediate parent MO reference(s) that the subassembly feeds (e.g. WH/MO/00020 or WH/MO/00020, WH/MO/00021) into the origin field ('Source' in the UI).\n* In the end, the lineage must be SO -> MO -> (Subassembly MO if needed) -> PO or SO -> PO.\n* You must sell this product at List Price.\n* On sales orders, set the commitment date.\n* On manufacturing orders, you must set the start date and the due date.\n* If you choose to manufacture, you must procure the components that are not in stock.\n* On purchase orders, you must set the delivery date.\n* Before releasing anything, read the Internal Notes/comments on stock, customers, vendors, and workcenters.\n\nCapacity constraints:\n- Treat workcenter capacity as a hard horizon-wide limit across all products that share the center.\n- Check each workcenter's Internal Notes in Odoo for the exact horizon-wide minute limit.\n- Assign workcenter on each manufacturing work order.\n- For each finished-goods supplier offer, respect min/max quantities as horizon-wide totals.\n- For each component supplier offer, respect min/max quantities as horizon-wide totals.\n- Use one consolidated PO per supplier offer (do not split a single offer across multiple POs).\n- Check each vendor's Internal Notes in Odoo for maximum order quantity limits.\n\nWork in the `odoo` ERP and commit the plan there.\n", "rubric": {"checks": {"answer_checks": [{"expect": 15476.0, "field": "assembly_cost", "tol_abs": 77.38, "type": "number"}, {"expect": 31, "field": "orders_accepted", "tol_abs": 0, "type": "number"}, {"expect": 0, "field": "orders_rejected", "tol_abs": 0, "type": "number"}, {"expect": 1009, "field": "units_purchased", "tol_abs": 0.01, "type": "number"}, {"expect": 636, "field": "units_manufactured", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": 713, "name": "confirmed_sale_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale'", "tol_abs": 0.01, "type": "sql"}, {"expect": 1009, "name": "confirmed_purchase_units", "sql": "SELECT COALESCE(SUM(l.qty),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_abs": 0.01, "type": "sql"}, {"expect": 636, "name": "confirmed_manufacture_units", "sql": "SELECT COALESCE(SUM(qty),0) FROM erpb_manufacturing_orders WHERE state='confirmed'", "tol_abs": 0.01, "type": "sql"}, {"expect": 0, "name": "mo_workcenters_exist", "sql": "SELECT COUNT(*) FROM erpb_manufacturing_orders m WHERE m.state='confirmed' AND NOT EXISTS (SELECT 1 FROM erpb_workcenters w WHERE w.code=m.workcenter_code)", "type": "sql"}, {"expect": 0, "name": "all_lines_at_list_price", "sql": "SELECT COUNT(*) FROM erpb_sale_order_lines l JOIN erpb_sale_orders o ON o.name=l.order_name WHERE o.state='sale' AND ABS(l.price_unit-4143.04)>0.01", "type": "sql"}, {"expect": 0, "name": "po_origin_links_to_so", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders p WHERE p.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_sale_orders s WHERE s.state='sale' AND p.origin LIKE '%'||s.name||'%')", "type": "sql"}, {"expect": 0, "name": "sale_commitment_dates_set", "sql": "SELECT COUNT(*) FROM erpb_sale_orders WHERE state='sale' AND (commitment_date IS NULL OR commitment_date='')", "type": "sql"}, {"expect": 0, "name": "purchase_dates_set", "sql": "SELECT COUNT(*) FROM erpb_purchase_orders WHERE state='purchase' AND (date_planned IS NULL OR date_planned='')", "type": "sql"}, {"expect": 0, "name": "po_lines_match_a_real_offer", "sql": "SELECT COUNT(*) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase' AND NOT EXISTS (SELECT 1 FROM erpb_vendor_offers v WHERE v.partner_ref=o.partner_ref AND v.product_code=l.product_code)", "type": "sql"}, {"expect": 670485.44, "name": "purchase_spend_matches_optimal", "sql": "SELECT COALESCE(SUM(l.qty*l.price_unit),0) FROM erpb_purchase_order_lines l JOIN erpb_purchase_orders o ON o.name=l.order_name WHERE o.state='purchase'", "tol_rel": 0.005, "type": "sql"}, {"tables": ["answers", "erpb_bom_components", "erpb_boms", "erpb_demand", "erpb_manufacturing_orders", "erpb_partners", "erpb_products", "erpb_purchase_order_lines", "erpb_purchase_orders", "erpb_sale_order_lines", "erpb_sale_orders", "erpb_stock", "erpb_vendor_offers", "erpb_workcenters"], "type": "writes_only"}], "trace_checks": [{"servers": ["odoo"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-056-2170-hard-17-shared-component-subassemblies-branch-assigned", "task_name": "erpbench/2170-hard-17-shared-component-subassemblies-branch-assigned", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-057-policy-validation-feb/documents/policy--travel-perdiem-and-substantiation.md", "task_files/lgr100-057-policy-validation-feb/documents/reference--gsa-mie-rate-table.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"conference_mie_rate": 92, "nonaccountable_amount": 612.4, "perdiem_due": 404.0, "reimbursable_total": 2932.5, "structural_exception_line": "EXP-2058", "structural_exception_reason": "A. Okonkwo's Chicago lodging is inside every per-line limit - receipt attached, business purpose stated, 306.20 a night against a 350 cap - but it was incurred 2025-12-08 and first substantiated 2026-02-19, 73 days later, outside the 60 day fixed-date safe harbour, so it falls out of the accountable plan and is paid through payroll as W-2 wages rather than reimbursed tax-free."}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "expense_audit", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-057-policy-validation-feb", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of TheAgentCompany research/external/repos/TheAgentCompany/workspaces/tasks/finance-expense-validation (expense extract + a rules file, find the mis-stated lines, report a per-category rollup). Kept: the spreadsheet-plus-policy surface and the rollup deliverable. Changed: the source grades keyword-to-category matching over 1,000 synthetic rows, which is arithmetic volume, so the port replaces the volume with one judgement — every one of the 13 lines passes the per-line test and the file is still wrong. Policy limits are grounded in research/external/articles/te-expense-audit--{gsa-mie-breakdown,gsa-perdiem-faq,ecfr-accountable-plan,irs}.md (GSA CONUS M&IE tiers and the 75% first/last-day rule, the 12-hour one-day gate, work-location locality rule, provided-meal exceptions, and the 26 CFR 1.62-2 60-day substantiation safe harbour with W-2 consequence). Differs from expense_audit/te-sample-feb, which tests per-line limits (meal cap, business class, missing receipt) and grades a violation count — here no line breaks a limit; and from expense_audit/threshold-shaving-h1, which grades a cross-period behavioural pattern over compliant claims — here the finding is a single line's date arithmetic and its payroll consequence. docs/HARD-LAYER-DESIGN.md M3 (non-collapse), M4 (money re-derivation off the rate table, never the claimed figure), M5 (the provided meal and the venue city live only in email, not in the extract).", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "email", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dev Bhatt · Payroll & Expense Lead · Teams 09:20**\n\nMorning — I need the February expense file signed off before the 5 March payroll cut-off and\nI'd rather you did it than me.\n\nPeople Ops dropped the extract on the shared drive. Everything on it looks clean: receipts are\nattached, nothing's over a cap, nobody flew business. So the per-line check tells me nothing,\nand I don't trust the per diem figures at all — those are whatever the traveller typed into the\nform, and we reprice them ourselves now under FIN-POL-022.\n\nThe other thing I got burned on in Q4: external audit picked up a reimbursement we paid out\ntax-free that shouldn't have been, and it cost me a W-2c. So before you send me a number, look\nat the dates and not just the amounts.\n", "rubric": {"checks": {"answer_checks": [{"expect": 404.0, "field": "perdiem_due", "tol_abs": 0.02, "type": "number"}, {"expect": 92, "field": "conference_mie_rate", "tol_abs": 0.01, "type": "number"}, {"expect": 612.4, "field": "nonaccountable_amount", "gt_sql": "SELECT ROUND(SUM(json_extract(cells,'$[6]')),2) FROM sheet_rows WHERE file='expense-extract-2026-02.xlsx' AND row_no>1 AND julianday(json_extract(cells,'$[8]')) - julianday(json_extract(cells,'$[3]')) > 60", "tol_abs": 0.02, "type": "number"}, {"expect": 2932.5, "field": "reimbursable_total", "tol_abs": 0.05, "type": "number"}, {"expect": ["EXP-2058"], "field": "structural_exception_line", "forbid": ["EXP-2041", "EXP-2042", "EXP-2043", "EXP-2044", "EXP-2049", "EXP-2050", "EXP-2051", "EXP-2052", "EXP-2055", "EXP-2056", "EXP-2057", "EXP-2059"], "type": "contains_all"}, {"expect": ["60"], "field": "structural_exception_reason", "forbid": ["duplicate", "missing receipt", "no receipt", "over the cap", "exceeds the cap", "business class"], "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs", "email"], "type": "required_servers"}, {"n": 2, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-057-policy-validation-feb", "task_name": "expense_audit/policy-validation-feb", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-058-te-sample-feb/documents/policy--travel-and-expense.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"out_of_policy_amount": 1512.4, "violating_report_ids": "EXP-1042, EXP-1047, EXP-1051", "violation_count": 3, "worst_violation_rule": "business class airfare requires VP pre-approval; economy is the standard"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "expense_audit", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-058-te-sample-feb", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#11 (T&E policy compliance sampling; Ramp-class agents claim >99% automation here)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 16:15**\n\nFebruary expenses — anything out of policy? I need the exceptions and what they cost us.\n", "rubric": {"checks": {"answer_checks": [{"expect": 3, "field": "violation_count", "tol_abs": 0, "type": "number"}, {"expect": ["EXP-1042", "EXP-1047", "EXP-1051"], "field": "violating_report_ids", "type": "contains_all"}, {"expect": 1512.4, "field": "out_of_policy_amount", "tol_abs": 0.02, "type": "number"}, {"expect": ["business class"], "field": "worst_violation_rule", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-058-te-sample-feb", "task_name": "expense_audit/te-sample-feb", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-059-te-sample-feb-esc-burie-quiet/documents/policy--travel-and-expense.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"out_of_policy_amount": 1512.4, "violating_report_ids": "EXP-1042, EXP-1047, EXP-1051", "violation_count": 3, "worst_violation_rule": "business class airfare requires VP pre-approval; economy is the standard"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "expense_audit", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-059-te-sample-feb-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of expense_audit/te-sample-feb via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#11 (T&E policy compliance sampling; Ramp-class agents claim >99% automation here)", "provenance": "variant", "version": "1.0.0", "walk_len": 4, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 16:15**\n\nFebruary expenses — anything out of policy? I need the exceptions and what they cost us.\n", "rubric": {"checks": {"answer_checks": [{"expect": 3, "field": "violation_count", "tol_abs": 0, "type": "number"}, {"expect": ["EXP-1042", "EXP-1047", "EXP-1051"], "field": "violating_report_ids", "type": "contains_all"}, {"expect": 1512.4, "field": "out_of_policy_amount", "tol_abs": 0.02, "type": "number"}, {"expect": ["business class"], "field": "worst_violation_rule", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-059-te-sample-feb-esc-burie-quiet", "task_name": "expense_audit/te-sample-feb-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-060-threshold-shaving-h1/documents/policy--expense-audit-detectors.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"detector_triggered": "yes", "shaving_claim_count": 6, "shaving_employee": "K. Osei", "shaving_total": 2850.74}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "expense_audit", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-060-threshold-shaving-h1", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/external/articles: Oracle Expenses threshold-shaving detector — claims in the 450-499 band against a 500 receipt threshold, more than 5 times in 6 months", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 11:55**\n\nTime for the periodic expense detectors — last six months. Every claim passed on its own, but run the patterns and tell me if anyone flags.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["osei"], "field": "shaving_employee", "type": "contains_all"}, {"expect": 6, "field": "shaving_claim_count", "tol_abs": 0, "type": "number"}, {"expect": 2850.74, "field": "shaving_total", "tol_abs": 0.02, "type": "number"}, {"expect": "yes", "field": "detector_triggered", "type": "yes_no"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-060-threshold-shaving-h1", "task_name": "expense_audit/threshold-shaving-h1", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"operating_margin_pct": 9.87, "revenue": 71043, "revenue_scale": "millions", "revenue_usd": 71043000000, "source_form": "10-K"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-061-scale-trap-lmt", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "closes tracked coverage gap tatqa.scale_unit (sim/coverage.py); NExTplusplus/TAT-QA grades `scale` (thousand/million/billion/percent) as a separately annotated field — research/evals-and-benchmarks.md; round-2 ledger row 28 (add a scale/unit field rather than widening tolerances until units stop mattering)", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 15:40**\n\nCounterparty pack for Lockheed Martin — I need FY2024 total net sales, and the operating\nmargin off the same filing.\n\nPut the units on it explicitly this time. Last pack went out with a revenue figure copied\nstraight off the statement table and no unit next to it, and it got read as dollars in the\ncommittee. Not repeating that.\n", "rubric": {"checks": {"answer_checks": [{"expect": 71043, "field": "revenue", "tol_rel": 0.001, "type": "number"}, {"expect": "millions", "field": "revenue_scale", "type": "scale"}, {"expect": 71043000000, "field": "revenue_usd", "tol_rel": 0.001, "type": "number"}, {"expect": 9.8715, "field": "operating_margin_pct", "tol_abs": 0.02, "type": "number"}, {"expect": ["10-K"], "field": "source_form", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-061-scale-trap-lmt", "task_name": "finance_qa/scale-trap-lmt", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-062-tac-10k-income-report/documents/policy--external-reporting-basis.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"cost_of_revenue": 642000000, "gross_profit": 840000000, "net_income": 412000000, "operating_income": 313000000, "operating_margin_pct": 21.12, "revenue": 1482000000, "revenue_growth_pct": 14.35, "source_form": "10-K for fiscal year 2025, accession 0009900413-26-000007, Item 8"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-062-tac-10k-income-report", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-create-10k-income-report (extract income-statement figures from 10-K Item 8 for multiple years, compile a structured table, and derive the year-over-year percentage change). The source's plumbing does not port: its ownCloud folder-navigation checkpoint, its written /workspace/net_income_report.csv, and its LLM-judged RocketChat message to David Wong are all environment-specific, and its checkpoint 4 graded an exact CSV cell set. Here the workbook becomes the frozen filings server (filings_companies/filings_facts/filings_documents, seeded as tasks/finance_qa/scale-trap-lmt does), the column structure becomes submit_answer fields, the CSV-shape checkpoint becomes the required_servers/reads_before_submit trace checks, and the derived percentage change is graded as revenue_growth_pct alongside a derived margin. The source's own trap - 'net income ... in section/item 8' and 'the name of registrant as specified in its charter' - is kept and sharpened: docs/HARD-LAYER-DESIGN.md M3 (non-collapse), M4 (the margin and the growth rate are recomputed from revenue and the line items, never read off a stated figure) and M5 (the governing basis lives only in FRP-02)", "provenance": "ported", "version": "1.0.0", "walk_len": 13, "walk_servers": ["docs", "filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Marta Ilyin · FP&A Manager · Teams 09:20**\n\nCalderwood Systems filed their FY2025 annual report last week. They're the comp our board pack\nbenchmarks us against, so I need their income statement summary rebuilt for the comp table on\nThursday, per FRP-02.\n\nTwo things I'd rather not repeat from last quarter. The margin we showed for them was about\nthree points better than the one in their accounts — the committee asked where it came from and\nI had nothing. And the growth line had them shrinking, which nobody in the room believed; it\nturned out we'd measured this year against a prior-year number they no longer report on that\nbasis.\n\nSo: their numbers, off the filing they actually stand behind, and tell me which filing that is.\n", "rubric": {"checks": {"answer_checks": [{"expect": 1482000000, "field": "revenue", "gt_sql": "SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='Revenues' AND fy='2025' AND period_end='2025-12-31'", "tol_rel": 0.0005, "type": "number"}, {"expect": 642000000, "field": "cost_of_revenue", "gt_sql": "SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='CostOfRevenue' AND fy='2025' AND period_end='2025-12-31'", "tol_rel": 0.0005, "type": "number"}, {"expect": 840000000, "field": "gross_profit", "gt_sql": "SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='GrossProfit' AND fy='2025' AND period_end='2025-12-31'", "tol_rel": 0.0005, "type": "number"}, {"expect": 313000000, "field": "operating_income", "gt_sql": "SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='OperatingIncomeLoss' AND fy='2025' AND period_end='2025-12-31'", "tol_rel": 0.0005, "type": "number"}, {"expect": 412000000, "field": "net_income", "gt_sql": "SELECT (SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='ProfitLoss' AND fy='2025' AND period_end='2025-12-31') - (SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='NetIncomeLossAttributableToNoncontrollingInterest' AND fy='2025' AND period_end='2025-12-31')", "tol_rel": 0.0005, "type": "number"}, {"expect": 21.1201, "field": "operating_margin_pct", "gt_sql": "SELECT ROUND(100.0*(SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='OperatingIncomeLoss' AND fy='2025' AND period_end='2025-12-31')/(SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='Revenues' AND fy='2025' AND period_end='2025-12-31'),4)", "tol_abs": 0.06, "type": "number"}, {"expect": 14.3519, "field": "revenue_growth_pct", "gt_sql": "SELECT ROUND(100.0*((SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='Revenues' AND fy='2025' AND period_end='2025-12-31')/(SELECT value FROM filings_facts WHERE cik='0009900413' AND concept='Revenues' AND fy='2025' AND period_end='2024-12-31')-1),4)", "tol_abs": 0.06, "type": "number"}, {"expect": ["10-K"], "field": "source_form", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings", "docs"], "type": "required_servers"}, {"n": 4, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-062-tac-10k-income-report", "task_name": "finance_qa/tac-10k-income-report", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"operating_income": 1624000000, "operating_margin_pct": 5.78, "quarter_revenue": 28095000000}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-063-tsla-operating-margin-q3", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark finance_qa (quarterly + format-constrained segment); facts captured from data.sec.gov 2026-08-10", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 15:11**\n\nWhat was Tesla's GAAP operating margin in the September 2025 quarter? Percent to two decimals, plus the two inputs.\n", "rubric": {"checks": {"answer_checks": [{"expect": 5.78, "field": "operating_margin_pct", "tol_abs": 0.02, "type": "number"}, {"expect": 1624000000, "field": "operating_income", "tol_rel": 0.001, "type": "number"}, {"expect": 28095000000, "field": "quarter_revenue", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-063-tsla-operating-margin-q3", "task_name": "finance_qa/tsla-operating-margin-q3", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"operating_income": 1624000000, "operating_margin_pct": 5.78, "quarter_revenue": 28095000000}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-064-tsla-operating-margin-q3-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of finance_qa/tsla-operating-margin-q3 via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark finance_qa (quarterly + format-constrained segment); facts captured from data.sec.gov 2026-08-10", "provenance": "variant", "version": "1.0.0", "walk_len": 4, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 15:11**\n\nWhat was Tesla's GAAP operating margin in the September 2025 quarter? Percent to two decimals, plus the two inputs.\n", "rubric": {"checks": {"answer_checks": [{"expect": 5.78, "field": "operating_margin_pct", "tol_abs": 0.02, "type": "number"}, {"expect": 1624000000, "field": "operating_income", "tol_rel": 0.001, "type": "number"}, {"expect": 28095000000, "field": "quarter_revenue", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-064-tsla-operating-margin-q3-esc-burie-quiet", "task_name": "finance_qa/tsla-operating-margin-q3-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"concepts_checked": "ResearchAndDevelopmentExpense (absent); snapshot holds only CostOfRevenue and InventoryNet for WMT", "figure_available": "no", "rd_expense": "none"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-065-unavailable-concept", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "FinanceBench pattern found in repo sweep: refusal scored separately from a wrong answer (research/external/repos/INDEX.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 13:33**\n\nCommittee asked for Walmart's R&D spend for the year ended 31 January 2025. Can you pull it? Tell me what you checked.\n", "rubric": {"checks": {"answer_checks": [{"expect": "no", "field": "figure_available", "type": "yes_no"}, {"field": "rd_expense", "type": "none_answer"}, {"expect": ["inventory"], "field": "concepts_checked", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-065-unavailable-concept", "task_name": "finance_qa/unavailable-concept", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"average_inventory": 55663500000, "cost_of_revenue": 511753000000, "inventory_turnover": 9.19}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-066-wmt-inventory-turnover", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark finance_qa (ratio segment); facts captured from data.sec.gov 2026-08-10", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 14:05**\n\nCommittee wants Walmart's inventory turnover for the year ended 31 January 2025, to two decimals. Show me the inputs you used.\n", "rubric": {"checks": {"answer_checks": [{"expect": 9.1937, "field": "inventory_turnover", "tol_abs": 0.02, "type": "number"}, {"expect": 511753000000, "field": "cost_of_revenue", "tol_rel": 0.001, "type": "number"}, {"expect": 55663500000, "field": "average_inventory", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-066-wmt-inventory-turnover", "task_name": "finance_qa/wmt-inventory-turnover", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"average_inventory": 55663500000, "cost_of_revenue": 511753000000, "inventory_turnover": 9.19}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-067-wmt-inventory-turnover-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of finance_qa/wmt-inventory-turnover via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark finance_qa (ratio segment); facts captured from data.sec.gov 2026-08-10", "provenance": "variant", "version": "1.0.0", "walk_len": 5, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 14:05**\n\nCommittee wants Walmart's inventory turnover for the year ended 31 January 2025, to two decimals. Show me the inputs you used.\n", "rubric": {"checks": {"answer_checks": [{"expect": 9.1937, "field": "inventory_turnover", "tol_abs": 0.02, "type": "number"}, {"expect": 511753000000, "field": "cost_of_revenue", "tol_rel": 0.001, "type": "number"}, {"expect": 55663500000, "field": "average_inventory", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-067-wmt-inventory-turnover-esc-burie-quiet", "task_name": "finance_qa/wmt-inventory-turnover-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"cat_current_assets": 45682000000, "difference": 46308000000, "larger_company": "ExxonMobil", "xom_current_assets": 91990000000}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-068-xom-cat-liquidity-compare", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark finance_qa (multi-company composite segment); exercises get_xbrl_frames across companies", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 09:48**\n\nScreening Exxon and Caterpillar side by side. At FY2024, who's carrying more current assets and by how much?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["Exxon"], "field": "larger_company", "forbid": ["Caterpillar"], "type": "contains_all"}, {"expect": 91990000000, "field": "xom_current_assets", "tol_rel": 0.001, "type": "number"}, {"expect": 45682000000, "field": "cat_current_assets", "tol_rel": 0.001, "type": "number"}, {"expect": 46308000000, "field": "difference", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"n": 2, "server": "filings", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-068-xom-cat-liquidity-compare", "task_name": "finance_qa/xom-cat-liquidity-compare", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"current_assets_usd": 91990000000, "period_end": "2024-12-31", "source_form": "10-K"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "easy", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-069-xom-current-assets", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "financebenchmark:finance_qa (user-brief example); GT = real XBRL fact captured from data.sec.gov 2026-08-10", "provenance": "ported", "version": "1.0.0", "walk_len": 3, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 11:20**\n\nFor the counterparty pack: ExxonMobil's total current assets at FY2024 year end. Tell me where it comes from so I can footnote it.\n", "rubric": {"checks": {"answer_checks": [{"expect": 91990000000, "field": "current_assets_usd", "tol_rel": 0.001, "type": "number"}, {"expect": ["2024-12-31"], "field": "period_end", "type": "contains_all"}, {"expect": ["10-K"], "field": "source_form", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-069-xom-current-assets", "task_name": "finance_qa/xom-current-assets", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"current_assets_usd": 91990000000, "period_end": "2024-12-31", "source_form": "10-K"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "easy", "family": "finance_qa", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-070-xom-current-assets-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of finance_qa/xom-current-assets via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: financebenchmark:finance_qa (user-brief example); GT = real XBRL fact captured from data.sec.gov 2026-08-10", "provenance": "variant", "version": "1.0.0", "walk_len": 3, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams 11:20**\n\nFor the counterparty pack: ExxonMobil's total current assets at FY2024 year end. Tell me where it comes from so I can footnote it.\n", "rubric": {"checks": {"answer_checks": [{"expect": 91990000000, "field": "current_assets_usd", "tol_rel": 0.001, "type": "number"}, {"expect": ["2024-12-31"], "field": "period_end", "type": "contains_all"}, {"expect": ["10-K"], "field": "source_form", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-070-xom-current-assets-esc-burie-quiet", "task_name": "finance_qa/xom-current-assets-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"debt_to_equity": 10.6105, "stockholders_equity": 344758000000.0, "total_liabilities": 3658056000000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-071-jpm-debt-to-equity-127", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark finance_qa item finance_qa-127; question verbatim, ground truth recomputed from the frozen SEC XBRL snapshot (debt-to-equity = Liabilities/StockholdersEquity @ 2024-12-31 (10-K)). FB grades this with an LLM judge; this clone pins the figure, its period and its source form as separate deterministic checks (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams**\n\nBenchmarking JPMorgan against peers for counterparty risk. What was JPMorgan Chase & Co.’s debt-to-equity ratio as of December 31, 2024?\n", "rubric": {"checks": {"answer_checks": [{"expect": 10.6105, "field": "debt_to_equity", "tol_abs": 0.02, "type": "number"}, {"expect": 3658056000000.0, "field": "total_liabilities", "tol_rel": 0.001, "type": "number"}, {"expect": 344758000000.0, "field": "stockholders_equity", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-071-jpm-debt-to-equity-127", "task_name": "finance_qa_fb/jpm-debt-to-equity-127", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"operating_income": 13762000000.0, "operating_margin_pct": 28.71, "revenue": 47941000000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "finance_qa_fb", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-072-ko-operating-margin-pct-035", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "clone of microsoft/FinanceBenchmark finance_qa item finance_qa-035; question verbatim, ground truth recomputed from the frozen SEC XBRL snapshot (operating margin = OperatingIncomeLoss/Revenues @ 2025-12-31 (10-K)). FB grades this with an LLM judge; this clone pins the figure, its period and its source form as separate deterministic checks (docs/PARITY.md)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["filings", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Dana Kim · Credit Manager · Teams**\n\nI'm updating margin assumptions for our FY2025 beverage forecast. What was Coca-Cola's operating margin for fiscal year 2024? Need it to calibrate our baseline.\n", "rubric": {"checks": {"answer_checks": [{"expect": 28.71, "field": "operating_margin_pct", "tol_abs": 0.02, "type": "number"}, {"expect": 13762000000.0, "field": "operating_income", "tol_rel": 0.001, "type": "number"}, {"expect": 47941000000.0, "field": "revenue", "tol_rel": 0.001, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["filings"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-072-ko-operating-margin-pct-035", "task_name": "finance_qa_fb/ko-operating-margin-pct-035", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-073-budget-variance-feb/documents/policy--fpna-variance-review.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"flagged_count": 4, "flagged_ids": "Engineering_Software_202602, Sales_Travel_202602, Marketing_Training_202602, Operations_Equipment_202601", "unfavourable_total_usd": 38515.4}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "fpna", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-073-budget-variance-feb", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-budget-variance (task.md + evaluator.py: budget vs actual by department/category/month, variance amount and %, flag >10% AND >$5,000, 'adjust for identified seasonal spending variations', identifiers as Department_Category_YYYYMM); the source's vague seasonality clause is resolved here into FIN-FPA-04 s4; docs/HARD-LAYER-DESIGN.md M3 (dual-limb non-collapse) and M5 (the front-loaded list has no field in either workbook and lives only in the policy document)", "provenance": "ported", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Iris Novak · FP&A Lead · Teams 08:20, Monday 2 March**\n\nFebruary's closed and the operating review is Thursday. I need the exception page off the FY26\nbudget and the Jan–Feb actuals — both workbooks are on the finance drive under my name.\n\nLast cycle we walked in with thirty lines on that page and spent forty minutes arguing about\nthe January facilities charge instead of the things that had actually moved. Ops walked out not\nbelieving any of it. So: run it the way FIN-FPA-04 says to run it, including what it says\nabout the categories that don't bill the way they're budgeted. If the policy wouldn't flag it,\nI don't want it on my page — I'd rather defend a short list than explain a long one.\n\nIDs in the usual Department_Category_YYYYMM shape so I can paste them straight into the deck.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["Engineering_Software_202602", "Sales_Travel_202602", "Marketing_Training_202602", "Operations_Equipment_202601"], "field": "flagged_ids", "forbid": ["Engineering_Facilities_202601", "Operations_Facilities_202601", "Engineering_Facilities_202602", "Operations_Facilities_202602", "Sales_Software_202602", "Marketing_Equipment_202602", "HR_Training_202601", "Engineering_Salaries_202602", "Sales_Salaries_202601"], "type": "contains_all"}, {"expect": 4, "field": "flagged_count", "gt_sql": "SELECT COUNT(*) FROM sheet_rows a JOIN sheet_rows b ON b.file='FY26-opex-budget.xlsx' AND b.row_no>1 AND json_extract(b.cells,'$[0]')=json_extract(a.cells,'$[0]') AND json_extract(b.cells,'$[1]')=json_extract(a.cells,'$[1]') AND json_extract(b.cells,'$[2]')=json_extract(a.cells,'$[2]') WHERE a.file='FY26-actuals-jan-feb.xlsx' AND a.row_no>1 AND json_extract(a.cells,'$[1]')<>'Facilities' AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>5000.0 AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>0.10*json_extract(b.cells,'$[3]')", "tol_abs": 0, "type": "number"}, {"expect": 38515.4, "field": "unfavourable_total_usd", "gt_sql": "SELECT ROUND(SUM(json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]')),2) FROM sheet_rows a JOIN sheet_rows b ON b.file='FY26-opex-budget.xlsx' AND b.row_no>1 AND json_extract(b.cells,'$[0]')=json_extract(a.cells,'$[0]') AND json_extract(b.cells,'$[1]')=json_extract(a.cells,'$[1]') AND json_extract(b.cells,'$[2]')=json_extract(a.cells,'$[2]') WHERE a.file='FY26-actuals-jan-feb.xlsx' AND a.row_no>1 AND json_extract(a.cells,'$[1]')<>'Facilities' AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>5000.0 AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>0.10*json_extract(b.cells,'$[3]')", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-073-budget-variance-feb", "task_name": "fpna/budget-variance-feb", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-074-budget-variance-feb-esc-burie-quiet/documents/policy--fpna-variance-review.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"flagged_count": 4, "flagged_ids": "Engineering_Software_202602, Sales_Travel_202602, Marketing_Training_202602, Operations_Equipment_202601", "unfavourable_total_usd": 38515.4}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "fpna", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-074-budget-variance-feb-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of fpna/budget-variance-feb via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-budget-variance (task.md + evaluator.py: budget vs actual by department/category/month, variance amount and %, flag >10% AND >$5,000, 'adjust for identified seasonal spending variations', identifiers as Department_Category_YYYYMM); the source's vague seasonality clause is resolved here into FIN-FPA-04 s4; docs/HARD-LAYER-DESIGN.md M3 (dual-limb non-collapse) and M5 (the front-loaded list has no field in either workbook and lives only in the policy document)", "provenance": "variant", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Iris Novak · FP&A Lead · Teams 08:20, Monday 2 March**\n\nFebruary's closed and the operating review is Thursday. I need the exception page off the FY26\nbudget and the Jan–Feb actuals — both workbooks are on the finance drive under my name.\n\nLast cycle we walked in with thirty lines on that page and spent forty minutes arguing about\nthe January facilities charge instead of the things that had actually moved. Ops walked out not\nbelieving any of it. So: run it the way the governing policy says to run it, including what it says\nabout the categories that don't bill the way they're budgeted. If the policy wouldn't flag it,\nI don't want it on my page — I'd rather defend a short list than explain a long one.\n\nIDs in the usual Department_Category_YYYYMM shape so I can paste them straight into the deck.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["Engineering_Software_202602", "Sales_Travel_202602", "Marketing_Training_202602", "Operations_Equipment_202601"], "field": "flagged_ids", "forbid": ["Engineering_Facilities_202601", "Operations_Facilities_202601", "Engineering_Facilities_202602", "Operations_Facilities_202602", "Sales_Software_202602", "Marketing_Equipment_202602", "HR_Training_202601", "Engineering_Salaries_202602", "Sales_Salaries_202601"], "type": "contains_all"}, {"expect": 4, "field": "flagged_count", "gt_sql": "SELECT COUNT(*) FROM sheet_rows a JOIN sheet_rows b ON b.file='FY26-opex-budget.xlsx' AND b.row_no>1 AND json_extract(b.cells,'$[0]')=json_extract(a.cells,'$[0]') AND json_extract(b.cells,'$[1]')=json_extract(a.cells,'$[1]') AND json_extract(b.cells,'$[2]')=json_extract(a.cells,'$[2]') WHERE a.file='FY26-actuals-jan-feb.xlsx' AND a.row_no>1 AND json_extract(a.cells,'$[1]')<>'Facilities' AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>5000.0 AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>0.10*json_extract(b.cells,'$[3]')", "tol_abs": 0, "type": "number"}, {"expect": 38515.4, "field": "unfavourable_total_usd", "gt_sql": "SELECT ROUND(SUM(json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]')),2) FROM sheet_rows a JOIN sheet_rows b ON b.file='FY26-opex-budget.xlsx' AND b.row_no>1 AND json_extract(b.cells,'$[0]')=json_extract(a.cells,'$[0]') AND json_extract(b.cells,'$[1]')=json_extract(a.cells,'$[1]') AND json_extract(b.cells,'$[2]')=json_extract(a.cells,'$[2]') WHERE a.file='FY26-actuals-jan-feb.xlsx' AND a.row_no>1 AND json_extract(a.cells,'$[1]')<>'Facilities' AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>5000.0 AND (json_extract(a.cells,'$[3]')-json_extract(b.cells,'$[3]'))>0.10*json_extract(b.cells,'$[3]')", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-074-budget-variance-feb-esc-burie-quiet", "task_name": "fpna/budget-variance-feb-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-075-payroll-attendance-tieout/documents/sop--payroll-register-review.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"discrepancy_employee_ids": "E-1044, E-1045", "net_variance_usd": -79.0, "overpayment_usd": 129.0, "underpaid_employee": "E-1045 Aisha Bello", "underpayment_usd": 208.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "fpna", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-075-payroll-attendance-tieout", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-check-attendance-payroll (attendance x rate -> earnings, with the rate card held in a source separate from attendance); reframed from 'compute earnings' to 'tie out what payroll actually paid' per docs/HARD-LAYER-DESIGN.md M3 (objective non-collapse), M4 (money re-derived from hours x approved rate, never read back off the register) and M5 (the LWOP payability rule exists only in SOP-PAY-04, not as a field on either workbook)", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Joel Tan · FP&A Manager · Teams 09:12**\n\nArdent sent back the register for the 16–28 Feb half and I have to sign it off before the pay\nfile releases Thursday — and the labour accrual goes in off the same number.\n\nI ran my usual eyeball over it: hours total agrees to the timekeeping export, gross total looks\nsane, so on the face of it there's nothing in there. But Comp pushed through the January grade\nchanges late and I don't fully trust what the bureau has on its master file, and HR had at\nleast one leave block in the period. Can you do the proper tie-out per SOP-PAY-04 rather than\nmy eyeball version?\n\nTwo things I care about: I don't want to fund a correction for something that isn't actually\nwrong, and if anyone was short-paid I need the name today so it goes on the off-cycle run.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["E-1044", "E-1045"], "field": "discrepancy_employee_ids", "forbid": ["E-1041", "E-1042", "E-1043", "E-1046", "E-1047"], "type": "contains_all"}, {"expect": 129.0, "field": "overpayment_usd", "gt_sql": "SELECT ROUND(SUM(MAX(p.gross - o.owed, 0)),2) FROM (SELECT json_extract(cells,'$[0]') AS id, json_extract(cells,'$[4]') AS gross FROM sheet_rows WHERE file='payroll-register-2026-02-H2.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-1%') p JOIN (SELECT json_extract(a.cells,'$[0]') AS id, (json_extract(a.cells,'$[3]') + CASE WHEN json_extract(a.cells,'$[5]')='LWOP' THEN 0 ELSE json_extract(a.cells,'$[4]') END) * json_extract(r.cells,'$[3]') AS owed FROM sheet_rows a JOIN sheet_rows r ON json_extract(r.cells,'$[0]')=json_extract(a.cells,'$[0]') WHERE a.file='attendance-export-2026-02-H2.xlsx' AND r.file='comp-rate-schedule-2026.xlsx' AND json_extract(a.cells,'$[0]') LIKE 'E-1%') o ON p.id=o.id", "tol_abs": 0.05, "type": "number"}, {"expect": 208.0, "field": "underpayment_usd", "gt_sql": "SELECT ROUND(-SUM(MIN(p.gross - o.owed, 0)),2) FROM (SELECT json_extract(cells,'$[0]') AS id, json_extract(cells,'$[4]') AS gross FROM sheet_rows WHERE file='payroll-register-2026-02-H2.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-1%') p JOIN (SELECT json_extract(a.cells,'$[0]') AS id, (json_extract(a.cells,'$[3]') + CASE WHEN json_extract(a.cells,'$[5]')='LWOP' THEN 0 ELSE json_extract(a.cells,'$[4]') END) * json_extract(r.cells,'$[3]') AS owed FROM sheet_rows a JOIN sheet_rows r ON json_extract(r.cells,'$[0]')=json_extract(a.cells,'$[0]') WHERE a.file='attendance-export-2026-02-H2.xlsx' AND r.file='comp-rate-schedule-2026.xlsx' AND json_extract(a.cells,'$[0]') LIKE 'E-1%') o ON p.id=o.id", "tol_abs": 0.05, "type": "number"}, {"expect": -79.0, "field": "net_variance_usd", "gt_sql": "SELECT ROUND(SUM(p.gross - o.owed),2) FROM (SELECT json_extract(cells,'$[0]') AS id, json_extract(cells,'$[4]') AS gross FROM sheet_rows WHERE file='payroll-register-2026-02-H2.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-1%') p JOIN (SELECT json_extract(a.cells,'$[0]') AS id, (json_extract(a.cells,'$[3]') + CASE WHEN json_extract(a.cells,'$[5]')='LWOP' THEN 0 ELSE json_extract(a.cells,'$[4]') END) * json_extract(r.cells,'$[3]') AS owed FROM sheet_rows a JOIN sheet_rows r ON json_extract(r.cells,'$[0]')=json_extract(a.cells,'$[0]') WHERE a.file='attendance-export-2026-02-H2.xlsx' AND r.file='comp-rate-schedule-2026.xlsx' AND json_extract(a.cells,'$[0]') LIKE 'E-1%') o ON p.id=o.id", "tol_abs": 0.05, "type": "number"}, {"expect": ["E-1045", "Aisha Bello"], "field": "underpaid_employee", "forbid": ["E-1041", "E-1042", "E-1043", "E-1044", "E-1046", "E-1047", "Devin Okafor", "Tomas Lindqvist", "Meera Raghunathan"], "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"n": 1, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-075-payroll-attendance-tieout", "task_name": "fpna/payroll-attendance-tieout", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-076-payroll-attendance-tieout-esc-burie-quiet/documents/sop--payroll-register-review.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"discrepancy_employee_ids": "E-1044, E-1045", "net_variance_usd": -79.0, "overpayment_usd": 129.0, "underpaid_employee": "E-1045 Aisha Bello", "underpayment_usd": 208.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "fpna", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-076-payroll-attendance-tieout-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of fpna/payroll-attendance-tieout via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-check-attendance-payroll (attendance x rate -> earnings, with the rate card held in a source separate from attendance); reframed from 'compute earnings' to 'tie out what payroll actually paid' per docs/HARD-LAYER-DESIGN.md M3 (objective non-collapse), M4 (money re-derived from hours x approved rate, never read back off the register) and M5 (the LWOP payability rule exists only in SOP-PAY-04, not as a field on either workbook)", "provenance": "variant", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Joel Tan · FP&A Manager · Teams 09:12**\n\nArdent sent back the register for the 16–28 Feb half and I have to sign it off before the pay\nfile releases Thursday — and the labour accrual goes in off the same number.\n\nI ran my usual eyeball over it: hours total agrees to the timekeeping export, gross total looks\nsane, so on the face of it there's nothing in there. But Comp pushed through the January grade\nchanges late and I don't fully trust what the bureau has on its master file, and HR had at\nleast one leave block in the period. Can you do the proper tie-out per the current policy rather than\nmy eyeball version?\n\nTwo things I care about: I don't want to fund a correction for something that isn't actually\nwrong, and if anyone was short-paid I need the name today so it goes on the off-cycle run.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["E-1044", "E-1045"], "field": "discrepancy_employee_ids", "forbid": ["E-1041", "E-1042", "E-1043", "E-1046", "E-1047"], "type": "contains_all"}, {"expect": 129.0, "field": "overpayment_usd", "gt_sql": "SELECT ROUND(SUM(MAX(p.gross - o.owed, 0)),2) FROM (SELECT json_extract(cells,'$[0]') AS id, json_extract(cells,'$[4]') AS gross FROM sheet_rows WHERE file='payroll-register-2026-02-H2.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-1%') p JOIN (SELECT json_extract(a.cells,'$[0]') AS id, (json_extract(a.cells,'$[3]') + CASE WHEN json_extract(a.cells,'$[5]')='LWOP' THEN 0 ELSE json_extract(a.cells,'$[4]') END) * json_extract(r.cells,'$[3]') AS owed FROM sheet_rows a JOIN sheet_rows r ON json_extract(r.cells,'$[0]')=json_extract(a.cells,'$[0]') WHERE a.file='attendance-export-2026-02-H2.xlsx' AND r.file='comp-rate-schedule-2026.xlsx' AND json_extract(a.cells,'$[0]') LIKE 'E-1%') o ON p.id=o.id", "tol_abs": 0.05, "type": "number"}, {"expect": 208.0, "field": "underpayment_usd", "gt_sql": "SELECT ROUND(-SUM(MIN(p.gross - o.owed, 0)),2) FROM (SELECT json_extract(cells,'$[0]') AS id, json_extract(cells,'$[4]') AS gross FROM sheet_rows WHERE file='payroll-register-2026-02-H2.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-1%') p JOIN (SELECT json_extract(a.cells,'$[0]') AS id, (json_extract(a.cells,'$[3]') + CASE WHEN json_extract(a.cells,'$[5]')='LWOP' THEN 0 ELSE json_extract(a.cells,'$[4]') END) * json_extract(r.cells,'$[3]') AS owed FROM sheet_rows a JOIN sheet_rows r ON json_extract(r.cells,'$[0]')=json_extract(a.cells,'$[0]') WHERE a.file='attendance-export-2026-02-H2.xlsx' AND r.file='comp-rate-schedule-2026.xlsx' AND json_extract(a.cells,'$[0]') LIKE 'E-1%') o ON p.id=o.id", "tol_abs": 0.05, "type": "number"}, {"expect": -79.0, "field": "net_variance_usd", "gt_sql": "SELECT ROUND(SUM(p.gross - o.owed),2) FROM (SELECT json_extract(cells,'$[0]') AS id, json_extract(cells,'$[4]') AS gross FROM sheet_rows WHERE file='payroll-register-2026-02-H2.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-1%') p JOIN (SELECT json_extract(a.cells,'$[0]') AS id, (json_extract(a.cells,'$[3]') + CASE WHEN json_extract(a.cells,'$[5]')='LWOP' THEN 0 ELSE json_extract(a.cells,'$[4]') END) * json_extract(r.cells,'$[3]') AS owed FROM sheet_rows a JOIN sheet_rows r ON json_extract(r.cells,'$[0]')=json_extract(a.cells,'$[0]') WHERE a.file='attendance-export-2026-02-H2.xlsx' AND r.file='comp-rate-schedule-2026.xlsx' AND json_extract(a.cells,'$[0]') LIKE 'E-1%') o ON p.id=o.id", "tol_abs": 0.05, "type": "number"}, {"expect": ["E-1045", "Aisha Bello"], "field": "underpaid_employee", "forbid": ["E-1041", "E-1042", "E-1043", "E-1044", "E-1046", "E-1047", "Devin Okafor", "Tomas Lindqvist", "Meera Raghunathan"], "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"n": 1, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-076-payroll-attendance-tieout-esc-burie-quiet", "task_name": "fpna/payroll-attendance-tieout-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-077-rd-tax-credit-asc/documents/rd-credit-policy.md", "task_files/lgr100-077-rd-tax-credit-asc/documents/tax-positions-2025.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"asc_credit": 61047.0, "base_amount": 345800.0, "base_period_years": "2022, 2023, 2024", "current_year_qre": 781850.0, "prior_three_year_qre_total": 2074800.0, "qualified_wages": 471200.0, "reduced_credit_elected": "no - the 2025 return positions memo declines the section 280C reduced credit, so Section B carries the full 14% credit"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "fpna", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-077-rd-tax-credit-asc", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-apply-tax-credit (IRS Form 6765 Section B alternative simplified credit built from a wage file plus a financials extract, with the ambiguous questions - section 280C election, existence of prior-3-year QREs - resolved by consulting a colleague). The judgement ports; the plumbing does not: the fillable PDF becomes graded answer fields through harness__submit_answer, the two workbooks become sheets/sheet_rows, the f6765 instructions become a docs policy (TAX-POL-11), and the RocketChat consultation with the finance director becomes the year's return positions memo. Escalated per docs/HARD-LAYER-DESIGN.md M3 (the base period is the objective the naive four-row average collapses), M4 (every component re-derived from line detail - wages x project time %, 65% contract-research haircut - never read off a totals row) and M5 (the qualification rules - US-sited research, quality-control testing, market research, depreciation, patent legal - exist only in TAX-POL-11, not as a field on either workbook)", "provenance": "ported", "version": "1.0.0", "walk_len": 8, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Renata Alvarez · Tax Manager · Teams 08:20**\n\nForm 6765 has to go into the return pack on Friday and the research credit is the only\ndiscrete item that moves our Q1 ETR, so I need Section B built properly today instead of the\nprovider's placeholder.\n\nTwo things to know about what you'll be working from. Tax has kept the QRE history tab since\nthe year the programme started, and the 2025 cost detail came straight off the R&D cost\ncentre with nobody filtering it — so please work it up from the line detail against TAX-POL-11\nrather than off any total already sitting on those tabs. Last year the provider took the tabs\nas they came and we ended up amending.\n\nCredit year is the tax year ended 31 December 2025. If the reduced-credit election bears on\nthe number, it's a settled position — don't re-litigate it, just read it.\n", "rubric": {"checks": {"answer_checks": [{"expect": 471200.0, "field": "qualified_wages", "gt_sql": "SELECT ROUND(SUM(json_extract(cells,'$[5]') * json_extract(cells,'$[6]') / 100.0),2) FROM sheet_rows WHERE file='rd-wage-detail-2025.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-3%' AND json_extract(cells,'$[3]')='qualified research' AND json_extract(cells,'$[4]') LIKE '%(US)%'", "tol_abs": 0.5, "type": "number"}, {"expect": 781850.0, "field": "current_year_qre", "gt_sql": "SELECT ROUND((SELECT SUM(json_extract(cells,'$[5]') * json_extract(cells,'$[6]') / 100.0) FROM sheet_rows WHERE file='rd-wage-detail-2025.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-3%' AND json_extract(cells,'$[3]')='qualified research' AND json_extract(cells,'$[4]') LIKE '%(US)%') + (SELECT SUM(CASE WHEN json_extract(cells,'$[1]')='contract research' THEN 0.65*json_extract(cells,'$[5]') ELSE json_extract(cells,'$[5]') END) FROM sheet_rows WHERE file='rd-nonwage-costs-2025.xlsx' AND json_extract(cells,'$[0]') LIKE 'C-%' AND json_extract(cells,'$[1]') IN ('supplies','computer rental','contract research') AND json_extract(cells,'$[4]') LIKE '%(US)%'),2)", "tol_abs": 0.5, "type": "number"}, {"expect": ["2022", "2024"], "field": "base_period_years", "type": "contains_all"}, {"expect": 2074800.0, "field": "prior_three_year_qre_total", "gt_sql": "SELECT ROUND(SUM(json_extract(cells,'$[2]')),2) FROM sheet_rows WHERE file='rd-qre-history.xlsx' AND CAST(json_extract(cells,'$[0]') AS INTEGER) BETWEEN 2022 AND 2024", "tol_abs": 0.5, "type": "number"}, {"expect": 345800.0, "field": "base_amount", "gt_sql": "SELECT ROUND(SUM(json_extract(cells,'$[2]'))/6.0,2) FROM sheet_rows WHERE file='rd-qre-history.xlsx' AND CAST(json_extract(cells,'$[0]') AS INTEGER) BETWEEN 2022 AND 2024", "tol_abs": 0.5, "type": "number"}, {"expect": 61047.0, "field": "asc_credit", "gt_sql": "SELECT ROUND(0.14 * ((SELECT SUM(json_extract(cells,'$[5]') * json_extract(cells,'$[6]') / 100.0) FROM sheet_rows WHERE file='rd-wage-detail-2025.xlsx' AND json_extract(cells,'$[0]') LIKE 'E-3%' AND json_extract(cells,'$[3]')='qualified research' AND json_extract(cells,'$[4]') LIKE '%(US)%') + (SELECT SUM(CASE WHEN json_extract(cells,'$[1]')='contract research' THEN 0.65*json_extract(cells,'$[5]') ELSE json_extract(cells,'$[5]') END) FROM sheet_rows WHERE file='rd-nonwage-costs-2025.xlsx' AND json_extract(cells,'$[0]') LIKE 'C-%' AND json_extract(cells,'$[1]') IN ('supplies','computer rental','contract research') AND json_extract(cells,'$[4]') LIKE '%(US)%') - (SELECT SUM(json_extract(cells,'$[2]'))/6.0 FROM sheet_rows WHERE file='rd-qre-history.xlsx' AND CAST(json_extract(cells,'$[0]') AS INTEGER) BETWEEN 2022 AND 2024)),2)", "tol_abs": 0.5, "type": "number"}, {"expect": "no", "field": "reduced_credit_elected", "type": "yes_no"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"n": 2, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-077-rd-tax-credit-asc", "task_name": "fpna/rd-tax-credit-asc", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-078-feb-saas-accrual/documents/accrual-sop.md", "task_files/lgr100-078-feb-saas-accrual/documents/cloudscale-msa-v1-superseded.md", "task_files/lgr100-078-feb-saas-accrual/documents/cloudscale-msa-v2.md"], "gold_output": {"expected_state_assertions": [{"expect": "draft", "name": "journal_staged_not_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": 83835.62, "name": "accrual_amount_rederived", "sql": "SELECT ROUND(total_debit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "journal_balances", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.005, "type": "sql"}, {"expect": "2026-02", "name": "posted_to_february", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": "600200", "name": "expense_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND debit>0", "type": "sql"}, {"expect": "210100", "name": "accrual_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND credit>0", "type": "sql"}, {"expect": "pending", "name": "approval_pending", "sql": "SELECT status FROM erp_approval_requests WHERE doc_type='Journal Entry' AND doc_id='GJ-00001'", "type": "sql"}, {"expect": 1, "name": "no_extra_journals", "sql": "SELECT COUNT(*) FROM erp_ledger_journals", "type": "sql"}], "submit_answer": {"accrual_amount": 83835.62, "journal_id": "GJ-00001", "journal_state": "draft", "next_step": "Pending Controller approval - approval request APR-00001 was raised because the journal exceeds the 25,000.00 staff-accountant delegation-of-authority threshold (policy DOA-JE-01). The Controller must approve it before it can be posted; it is not booked yet.", "service_days": 17}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "journal_entry", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-078-feb-saas-accrual", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "docs/HARD-LAYER-DESIGN.md M1/M3/M4/M5; DoA shape from ERPNext Authorization Rule (research/erp-mcp-tool-census.md); period lock from research/odoo-domain.md; accrual practice from research/external/articles/month-end-close--numeric.md", "provenance": "ported", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Controller · Teams 08:20**\n\nFebruary close — the CloudScale subscription needs accruing. They invoice annually in\narrears so nothing has come through AP for the platform itself, but we've had the service\nsince it cut over last month.\n\nRaise the accrual into February per SOP-GL-02 and take it as far as you can. Careful with\nthe paperwork on that one, they re-papered the agreement recently.\n\nTell me where it ends up — I need to know what's actually booked before I sign off the\nclose, not what's been keyed.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["GJ-00001"], "field": "journal_id", "type": "contains_all"}, {"expect": 83835.62, "field": "accrual_amount", "gt_sql": "SELECT ROUND(total_debit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": 17, "field": "service_days", "tol_abs": 0.01, "type": "number"}, {"expect": ["draft"], "field": "journal_state", "type": "contains_all"}, {"expect": ["approv"], "field": "next_step", "type": "contains_all"}], "state_checks": [{"expect": "draft", "name": "journal_staged_not_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": 83835.62, "name": "accrual_amount_rederived", "sql": "SELECT ROUND(total_debit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "journal_balances", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.005, "type": "sql"}, {"expect": "2026-02", "name": "posted_to_february", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": "600200", "name": "expense_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND debit>0", "type": "sql"}, {"expect": "210100", "name": "accrual_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND credit>0", "type": "sql"}, {"expect": "pending", "name": "approval_pending", "sql": "SELECT status FROM erp_approval_requests WHERE doc_type='Journal Entry' AND doc_id='GJ-00001'", "type": "sql"}, {"expect": 1, "name": "no_extra_journals", "sql": "SELECT COUNT(*) FROM erp_ledger_journals", "type": "sql"}, {"tables": ["answers", "erp_ledger_journals", "erp_ledger_journal_lines", "erp_approval_requests", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-078-feb-saas-accrual", "task_name": "journal_entry/feb-saas-accrual", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-079-feb-saas-accrual-esc-burie-quiet/documents/accrual-sop.md", "task_files/lgr100-079-feb-saas-accrual-esc-burie-quiet/documents/cloudscale-msa-v1-superseded.md", "task_files/lgr100-079-feb-saas-accrual-esc-burie-quiet/documents/cloudscale-msa-v2.md"], "gold_output": {"expected_state_assertions": [{"expect": "draft", "name": "journal_staged_not_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": 83835.62, "name": "accrual_amount_rederived", "sql": "SELECT ROUND(total_debit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "journal_balances", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.005, "type": "sql"}, {"expect": "2026-02", "name": "posted_to_february", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": "600200", "name": "expense_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND debit>0", "type": "sql"}, {"expect": "210100", "name": "accrual_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND credit>0", "type": "sql"}, {"expect": "pending", "name": "approval_pending", "sql": "SELECT status FROM erp_approval_requests WHERE doc_type='Journal Entry' AND doc_id='GJ-00001'", "type": "sql"}, {"expect": 1, "name": "no_extra_journals", "sql": "SELECT COUNT(*) FROM erp_ledger_journals", "type": "sql"}], "submit_answer": {"accrual_amount": 83835.62, "journal_id": "GJ-00001", "journal_state": "draft", "next_step": "Pending Controller approval - approval request APR-00001 was raised because the journal exceeds the 25,000.00 staff-accountant delegation-of-authority threshold (policy DOA-JE-01). The Controller must approve it before it can be posted; it is not booked yet.", "service_days": 17}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "journal_entry", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-079-feb-saas-accrual-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of journal_entry/feb-saas-accrual via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: docs/HARD-LAYER-DESIGN.md M1/M3/M4/M5; DoA shape from ERPNext Authorization Rule (research/erp-mcp-tool-census.md); period lock from research/odoo-domain.md; accrual practice from research/external/articles/month-end-close--numeric.md", "provenance": "variant", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · Controller · Teams 08:20**\n\nFebruary close — the CloudScale subscription needs accruing. They invoice annually in\narrears so nothing has come through AP for the platform itself, but we've had the service\nsince it cut over last month.\n\nRaise the accrual into February per the current policy and take it as far as you can. Careful with\nthe paperwork on that one, they re-papered the agreement recently.\n\nTell me where it ends up — I need to know what's actually booked before I sign off the\nclose, not what's been keyed.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["GJ-00001"], "field": "journal_id", "type": "contains_all"}, {"expect": 83835.62, "field": "accrual_amount", "gt_sql": "SELECT ROUND(total_debit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": 17, "field": "service_days", "tol_abs": 0.01, "type": "number"}, {"expect": ["draft"], "field": "journal_state", "type": "contains_all"}, {"expect": ["approv"], "field": "next_step", "type": "contains_all"}], "state_checks": [{"expect": "draft", "name": "journal_staged_not_posted", "sql": "SELECT state FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": 83835.62, "name": "accrual_amount_rederived", "sql": "SELECT ROUND(total_debit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "journal_balances", "sql": "SELECT ROUND(total_debit-total_credit,2) FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "tol_abs": 0.005, "type": "sql"}, {"expect": "2026-02", "name": "posted_to_february", "sql": "SELECT period_id FROM erp_ledger_journals WHERE journal_id='GJ-00001'", "type": "sql"}, {"expect": "600200", "name": "expense_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND debit>0", "type": "sql"}, {"expect": "210100", "name": "accrual_line_account", "sql": "SELECT account_code FROM erp_ledger_journal_lines WHERE journal_id='GJ-00001' AND credit>0", "type": "sql"}, {"expect": "pending", "name": "approval_pending", "sql": "SELECT status FROM erp_approval_requests WHERE doc_type='Journal Entry' AND doc_id='GJ-00001'", "type": "sql"}, {"expect": 1, "name": "no_extra_journals", "sql": "SELECT COUNT(*) FROM erp_ledger_journals", "type": "sql"}, {"tables": ["answers", "erp_ledger_journals", "erp_ledger_journal_lines", "erp_approval_requests", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-079-feb-saas-accrual-esc-burie-quiet", "task_name": "journal_entry/feb-saas-accrual-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-080-friday-run-mar06/documents/sop--payment-run.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"discount_captured": 437.11, "excluded_invoice": "PPINV-104", "exclusion_reason": "vendor SYNVEN-0044 is on payment hold", "net_payment_total": 36418.39, "proposal_invoices": "PPINV-101, PPINV-102, PPINV-103"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "payment_proposal", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-080-friday-run-mar06", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#3 (D365 payment-proposal automation: agent assembles, human posts); discount-capture policy AP-POL-007", "provenance": "ported", "version": "1.0.0", "walk_len": 7, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 09:10**\n\nBuild me Friday's payment proposal (pay date 6 March) for the PPINV batch. I'll review and post it — just tell me what's in, what's out, and what it costs us net.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PPINV-101", "PPINV-102", "PPINV-103"], "field": "proposal_invoices", "type": "contains_all"}, {"expect": ["PPINV-104"], "field": "excluded_invoice", "type": "contains_all"}, {"expect": ["hold"], "field": "exclusion_reason", "type": "contains_all"}, {"expect": 437.11, "field": "discount_captured", "tol_abs": 0.02, "type": "number"}, {"expect": 36418.39, "field": "net_payment_total", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-080-friday-run-mar06", "task_name": "payment_proposal/friday-run-mar06", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-081-friday-run-mar06-esc-burie-quiet/documents/sop--payment-run.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"discount_captured": 437.11, "excluded_invoice": "PPINV-104", "exclusion_reason": "vendor SYNVEN-0044 is on payment hold", "net_payment_total": 36418.39, "proposal_invoices": "PPINV-101, PPINV-102, PPINV-103"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "payment_proposal", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-081-friday-run-mar06-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of payment_proposal/friday-run-mar06 via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#3 (D365 payment-proposal automation: agent assembles, human posts); discount-capture policy AP-POL-007", "provenance": "variant", "version": "1.0.0", "walk_len": 7, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 09:10**\n\nBuild me Friday's payment proposal (pay date 6 March) for the PPINV batch. I'll review and post it — just tell me what's in, what's out, and what it costs us net.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PPINV-101", "PPINV-102", "PPINV-103"], "field": "proposal_invoices", "type": "contains_all"}, {"expect": ["PPINV-104"], "field": "excluded_invoice", "type": "contains_all"}, {"expect": ["hold"], "field": "exclusion_reason", "type": "contains_all"}, {"expect": 437.11, "field": "discount_captured", "tol_abs": 0.02, "type": "number"}, {"expect": 36418.39, "field": "net_payment_total", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-081-friday-run-mar06-esc-burie-quiet", "task_name": "payment_proposal/friday-run-mar06-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-082-shortfall-mar06/documents/payment-run-sop.md"], "gold_output": {"expected_state_assertions": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": "PINV-101,PINV-105,PINV-106", "name": "paid_set", "sql": "SELECT GROUP_CONCAT(invoice) FROM (SELECT invoice FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid' ORDER BY invoice)", "type": "sql"}, {"expect": 187800.0, "name": "paid_total_rederived", "sql": "SELECT ROUND(total_paid,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": "vendor_on_hold", "name": "hold_reason", "sql": "SELECT reason_code FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='PINV-103'", "type": "sql"}, {"expect": "disputed", "name": "dispute_reason", "sql": "SELECT reason_code FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='PINV-104'", "type": "sql"}, {"expect": "PINV-102,PINV-107", "name": "cash_deferrals", "sql": "SELECT GROUP_CONCAT(invoice) FROM (SELECT invoice FROM erp_payment_run_lines WHERE run_id='PR-00001' AND reason_code='insufficient_cash' ORDER BY invoice)", "type": "sql"}, {"expect": 0, "name": "no_obligation_left_undecided", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition NOT IN ('paid','rejected')", "type": "sql"}], "submit_answer": {"paid_invoices": "PINV-101, PINV-105, PINV-106", "run_id": "PR-00001", "shortfall": 117800.0, "total_paid": 187800.0, "unpaid_invoices": "PINV-102, PINV-103, PINV-104, PINV-107"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "payment_run", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-082-shortfall-mar06", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "docs/HARD-LAYER-DESIGN.md M1/M2/M3/M4/M5; unsat-demand mechanic from ERP-Bench (77/300 tasks tagged unsat_demand, research/erp-bench-deep-dive.md); discount-capture policy from research/domain-workflows.md §2; SAP F110 payment proposals per research/external/articles/payment-runs-early-pay-discount--sap-f110.md", "provenance": "ported", "version": "1.0.0", "walk_len": 11, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 07:58, Friday 6 March**\n\nMorning — I'm out at the bank review until this afternoon, can you build and commit today's\nrun for the PAYRUN vendor group? Operating account.\n\nCash is tight this week so it won't all go out. Follow SOP-AP-04 exactly — I don't want a\nrepeat of last month where two obligations just vanished off the proposal and nobody could\nsay why. Everything on the proposal needs to end up either paid or explicitly not paid with\na reason on it.\n\nCheck the AP mailbox before you build it. Something came in this week.\n\nCommit the run when you're done.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PR-00001"], "field": "run_id", "type": "contains_all"}, {"expect": 187800.0, "field": "total_paid", "gt_sql": "SELECT ROUND(SUM(net_amount),2) FROM erp_payment_run_lines WHERE disposition='paid'", "tol_abs": 0.05, "type": "number"}, {"expect": 117800.0, "field": "shortfall", "gt_sql": "SELECT ROUND(eligible_net - cash_available,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": ["PINV-101", "PINV-105", "PINV-106"], "field": "paid_invoices", "type": "contains_all"}, {"expect": ["PINV-102", "PINV-103", "PINV-104", "PINV-107"], "field": "unpaid_invoices", "type": "contains_all"}], "state_checks": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": "PINV-101,PINV-105,PINV-106", "name": "paid_set", "sql": "SELECT GROUP_CONCAT(invoice) FROM (SELECT invoice FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid' ORDER BY invoice)", "type": "sql"}, {"expect": 187800.0, "name": "paid_total_rederived", "sql": "SELECT ROUND(total_paid,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "sql"}, {"expect": "vendor_on_hold", "name": "hold_reason", "sql": "SELECT reason_code FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='PINV-103'", "type": "sql"}, {"expect": "disputed", "name": "dispute_reason", "sql": "SELECT reason_code FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='PINV-104'", "type": "sql"}, {"expect": "PINV-102,PINV-107", "name": "cash_deferrals", "sql": "SELECT GROUP_CONCAT(invoice) FROM (SELECT invoice FROM erp_payment_run_lines WHERE run_id='PR-00001' AND reason_code='insufficient_cash' ORDER BY invoice)", "type": "sql"}, {"expect": 0, "name": "no_obligation_left_undecided", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition NOT IN ('paid','rejected')", "type": "sql"}, {"tables": ["answers", "erp_payment_runs", "erp_payment_run_lines", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs", "email"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-082-shortfall-mar06", "task_name": "payment_run/shortfall-mar06", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-083-withholding-mar13/documents/withholding-policy.md"], "gold_output": {"expected_state_assertions": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": 12000.0, "name": "meridian_withheld_at_statutory_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-803'", "tol_abs": 0.05, "type": "sql"}, {"expect": 3000.0, "name": "treaty_vendor_withheld_at_reduced_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-804'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "certified_domestic_not_withheld", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-801'", "tol_abs": 0.005, "type": "sql"}, {"expect": 4, "name": "all_four_paid", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid'", "type": "sql"}], "submit_answer": {"meridian_rate_pct": 30, "run_id": "PR-00001", "total_gross": 140000.0, "total_paid": 117800.0, "total_withheld": 22200.0, "withheld_vendors": "WVEN-02, WVEN-03, WVEN-04"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "payment_run", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-083-withholding-mar13", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "closes tracked coverage gap erpnext.withholding_tax; Tax Withholding Category + certificate model from research/erp-mcp-tool-census.md (ERPNext); rates are US statutory (IRC 3406 backup 24%, IRC 1441 non-resident 30%, treaty-reduced 15%); docs/HARD-LAYER-DESIGN.md M1/M3/M4/M5", "provenance": "ported", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 08:15, Friday 13 March**\n\nContractor run today — vendor group CONTRACT, operating account. Four invoices, all due today,\ncash is fine.\n\nThese are the ones with withholding on them, so please work the tax position per SOP-AP-09\nbefore you commit. Meridian have emailed about their rate again; check it rather than take\nit, tax were quite pointed about that in January.\n\nEverything on the proposal gets paid — I just need the amounts right and the withholding\nbroken out.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PR-00001"], "field": "run_id", "type": "contains_all"}, {"expect": 140000.0, "field": "total_gross", "gt_sql": "SELECT ROUND(SUM(gross_amount),2) FROM erp_payment_run_lines WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": 22200.0, "field": "total_withheld", "gt_sql": "SELECT ROUND(SUM(withholding),2) FROM erp_payment_run_lines WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": 117800.0, "field": "total_paid", "gt_sql": "SELECT ROUND(total_paid,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": ["WVEN-02", "WVEN-03", "WVEN-04"], "field": "withheld_vendors", "forbid": ["WVEN-01"], "type": "contains_all"}, {"expect": 30, "field": "meridian_rate_pct", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": 12000.0, "name": "meridian_withheld_at_statutory_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-803'", "tol_abs": 0.05, "type": "sql"}, {"expect": 3000.0, "name": "treaty_vendor_withheld_at_reduced_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-804'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "certified_domestic_not_withheld", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-801'", "tol_abs": 0.005, "type": "sql"}, {"expect": 4, "name": "all_four_paid", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid'", "type": "sql"}, {"tables": ["answers", "erp_payment_runs", "erp_payment_run_lines", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-083-withholding-mar13", "task_name": "payment_run/withholding-mar13", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-084-withholding-mar13-esc-burie-quiet/documents/withholding-policy.md"], "gold_output": {"expected_state_assertions": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": 12000.0, "name": "meridian_withheld_at_statutory_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-803'", "tol_abs": 0.05, "type": "sql"}, {"expect": 3000.0, "name": "treaty_vendor_withheld_at_reduced_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-804'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "certified_domestic_not_withheld", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-801'", "tol_abs": 0.005, "type": "sql"}, {"expect": 4, "name": "all_four_paid", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid'", "type": "sql"}], "submit_answer": {"meridian_rate_pct": 30, "run_id": "PR-00001", "total_gross": 140000.0, "total_paid": 117800.0, "total_withheld": 22200.0, "withheld_vendors": "WVEN-02, WVEN-03, WVEN-04"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "payment_run", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-084-withholding-mar13-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of payment_run/withholding-mar13 via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: closes tracked coverage gap erpnext.withholding_tax; Tax Withholding Category + certificate model from research/erp-mcp-tool-census.md (ERPNext); rates are US statutory (IRC 3406 backup 24%, IRC 1441 non-resident 30%, treaty-reduced 15%); docs/HARD-LAYER-DESIGN.md M1/M3/M4/M5", "provenance": "variant", "version": "1.0.0", "walk_len": 10, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Robin Vale · Treasury · Teams 08:15, Friday 13 March**\n\nContractor run today — vendor group CONTRACT, operating account. Four invoices, all due today,\ncash is fine.\n\nThese are the ones with withholding on them, so please work the tax position per the current policy\nbefore you commit. Meridian have emailed about their rate again; check it rather than take\nit, tax were quite pointed about that in January.\n\nEverything on the proposal gets paid — I just need the amounts right and the withholding\nbroken out.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PR-00001"], "field": "run_id", "type": "contains_all"}, {"expect": 140000.0, "field": "total_gross", "gt_sql": "SELECT ROUND(SUM(gross_amount),2) FROM erp_payment_run_lines WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": 22200.0, "field": "total_withheld", "gt_sql": "SELECT ROUND(SUM(withholding),2) FROM erp_payment_run_lines WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": 117800.0, "field": "total_paid", "gt_sql": "SELECT ROUND(total_paid,2) FROM erp_payment_runs WHERE run_id='PR-00001'", "tol_abs": 0.05, "type": "number"}, {"expect": ["WVEN-02", "WVEN-03", "WVEN-04"], "field": "withheld_vendors", "forbid": ["WVEN-01"], "type": "contains_all"}, {"expect": 30, "field": "meridian_rate_pct", "tol_abs": 0.01, "type": "number"}], "state_checks": [{"expect": "committed", "name": "run_committed", "sql": "SELECT state FROM erp_payment_runs WHERE run_id='PR-00001'", "type": "sql"}, {"expect": 12000.0, "name": "meridian_withheld_at_statutory_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-803'", "tol_abs": 0.05, "type": "sql"}, {"expect": 3000.0, "name": "treaty_vendor_withheld_at_reduced_rate", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-804'", "tol_abs": 0.05, "type": "sql"}, {"expect": 0.0, "name": "certified_domestic_not_withheld", "sql": "SELECT ROUND(withholding,2) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND invoice='WINV-801'", "tol_abs": 0.005, "type": "sql"}, {"expect": 4, "name": "all_four_paid", "sql": "SELECT COUNT(*) FROM erp_payment_run_lines WHERE run_id='PR-00001' AND disposition='paid'", "type": "sql"}, {"tables": ["answers", "erp_payment_runs", "erp_payment_run_lines", "erp_confirm_tokens", "erp_audit_trail"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-084-withholding-mar13-esc-burie-quiet", "task_name": "payment_run/withholding-mar13-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-085-approval-evidence-q1/inputs/pbc-request-03.csv"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"inv201_approval_date": "2026-01-18", "inv201_approver": "Priya Shah", "inv201_evidence_location": "shared mailbox thread (approval reply from AP Manager)", "inv202_approver": "none", "inv202_finding": "No approval evidence located in the ERP, mailbox, or document library — control exception to report"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "pbc", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-085-approval-evidence-q1", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#9 + chaos pattern 'approvals live in email threads'", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · email, Mon 09:20**\n\nRowan & Field sent PBC request 3 — it's in your working folder. They want approval evidence on the two invoices listed. Give me what you can actually support.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["priya"], "field": "inv201_approver", "type": "contains_all"}, {"expect": ["2026-01-18"], "field": "inv201_approval_date", "type": "contains_all"}, {"expect": ["mail"], "field": "inv201_evidence_location", "type": "contains_all"}, {"field": "inv202_approver", "type": "none_answer"}, {"expect": ["no", "evidence"], "field": "inv202_finding", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "email"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-085-approval-evidence-q1", "task_name": "pbc/approval-evidence-q1", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-086-approval-evidence-q1-esc-burie-quiet/inputs/pbc-request-03.csv"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"inv201_approval_date": "2026-01-18", "inv201_approver": "Priya Shah", "inv201_evidence_location": "shared mailbox thread (approval reply from AP Manager)", "inv202_approver": "none", "inv202_finding": "No approval evidence located in the ERP, mailbox, or document library — control exception to report"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "pbc", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-086-approval-evidence-q1-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of pbc/approval-evidence-q1 via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#9 + chaos pattern 'approvals live in email threads'", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · email, Mon 09:20**\n\nRowan & Field sent PBC request 3 — it's in your working folder. They want approval evidence on the two invoices listed. Give me what you can actually support.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["priya"], "field": "inv201_approver", "type": "contains_all"}, {"expect": ["2026-01-18"], "field": "inv201_approval_date", "type": "contains_all"}, {"expect": ["mail"], "field": "inv201_evidence_location", "type": "contains_all"}, {"field": "inv202_approver", "type": "none_answer"}, {"expect": ["no", "evidence"], "field": "inv202_finding", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "email"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-086-approval-evidence-q1-esc-burie-quiet", "task_name": "pbc/approval-evidence-q1-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-087-sampling-projection-q1/documents/sop--audit-sampling-method.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"exceeds_materiality": "yes", "method_used": "ratio projection: sample misstatement x (population size / sample size) = 3,000 x 20", "population_size": 1000, "projected_misstatement": 60000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "pbc", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-087-sampling-projection-q1", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/external/articles: PCAOB AS 2315 projection arithmetic (50 of 1,000 sampled, $3,000 found -> $60,000 projected)", "provenance": "ported", "version": "1.0.0", "walk_len": 4, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 14:40**\n\nThe auditors sent their Q1 sample results. What does that imply for the population as a whole — are we in trouble?\n", "rubric": {"checks": {"answer_checks": [{"expect": 60000.0, "field": "projected_misstatement", "tol_abs": 0.02, "type": "number"}, {"expect": 1000, "field": "population_size", "tol_abs": 0, "type": "number"}, {"expect": "yes", "field": "exceeds_materiality", "type": "yes_no"}, {"expect": ["ratio"], "field": "method_used", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-087-sampling-projection-q1", "task_name": "pbc/sampling-projection-q1", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-088-sampling-projection-q1-esc-burie-quiet/documents/sop--audit-sampling-method.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"exceeds_materiality": "yes", "method_used": "ratio projection: sample misstatement x (population size / sample size) = 3,000 x 20", "population_size": 1000, "projected_misstatement": 60000.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "medium", "family": "pbc", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-088-sampling-projection-q1-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of pbc/sampling-projection-q1 via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/external/articles: PCAOB AS 2315 projection arithmetic (50 of 1,000 sampled, $3,000 found -> $60,000 projected)", "provenance": "variant", "version": "1.0.0", "walk_len": 4, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Sam Rivera · Controller · Teams 14:40**\n\nThe auditors sent their Q1 sample results. What does that imply for the population as a whole — are we in trouble?\n", "rubric": {"checks": {"answer_checks": [{"expect": 60000.0, "field": "projected_misstatement", "tol_abs": 0.02, "type": "number"}, {"expect": 1000, "field": "population_size", "tol_abs": 0, "type": "number"}, {"expect": "yes", "field": "exceeds_materiality", "type": "yes_no"}, {"expect": ["ratio"], "field": "method_used", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-088-sampling-projection-q1-esc-burie-quiet", "task_name": "pbc/sampling-projection-q1-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"clean_invoice": "TWINV-301", "price_variance_amount": 300.0, "price_variance_invoice": "TWINV-302", "qty_over_billed_units": 40, "qty_variance_invoice": "TWINV-303"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "threeway_match", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-089-ppinv-exceptions-mar", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#6 (unparked by PO/receipt schema); Oracle 26B GA agents do this workflow", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 10:25**\n\nThree invoices are stuck in the match queue — TWINV-301, 302 and 303. What's wrong with each, and quantify it.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["TWINV-301"], "field": "clean_invoice", "type": "contains_all"}, {"expect": ["TWINV-302"], "field": "price_variance_invoice", "type": "contains_all"}, {"expect": 300.0, "field": "price_variance_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["TWINV-303"], "field": "qty_variance_invoice", "type": "contains_all"}, {"expect": 40, "field": "qty_over_billed_units", "tol_abs": 0, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-089-ppinv-exceptions-mar", "task_name": "threeway_match/ppinv-exceptions-mar", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [], "submit_answer": {"clean_invoice": "TWINV-301", "price_variance_amount": 300.0, "price_variance_invoice": "TWINV-302", "qty_over_billed_units": 40, "qty_variance_invoice": "TWINV-303"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "threeway_match", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-090-ppinv-exceptions-mar-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of threeway_match/ppinv-exceptions-mar via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/finance-agent-workflows.md wf#6 (unparked by PO/receipt schema); Oracle 26B GA agents do this workflow", "provenance": "variant", "version": "1.0.0", "walk_len": 5, "walk_servers": ["erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 10:25**\n\nThree invoices are stuck in the match queue — TWINV-301, 302 and 303. What's wrong with each, and quantify it.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["TWINV-301"], "field": "clean_invoice", "type": "contains_all"}, {"expect": ["TWINV-302"], "field": "price_variance_invoice", "type": "contains_all"}, {"expect": 300.0, "field": "price_variance_amount", "tol_abs": 0.01, "type": "number"}, {"expect": ["TWINV-303"], "field": "qty_variance_invoice", "type": "contains_all"}, {"expect": 40, "field": "qty_over_billed_units", "tol_abs": 0, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp"], "type": "required_servers"}, {"n": 3, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-090-ppinv-exceptions-mar-esc-burie-quiet", "task_name": "threeway_match/ppinv-exceptions-mar-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-091-tac-invoice-matching/documents/sop--vendor-statement-reconciliation.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"applied_total_usd": 113197.75, "exception_invoices": "HF-2203, HF-2207, HF-2218, HF-2222", "net_variance_usd": 3240.5, "out_of_scope_payment": "PMT-8817, which references HF-2190 - not on this statement", "overpaid_usd": 2750.0, "statement_total_usd": 116438.25, "underpaid_usd": 5990.5}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "threeway_match", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-091-tac-invoice-matching", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-invoice-matching (match a payments file to a reference invoice set, handling split and combined payments, and identify the mismatches; its deliverable is a written flagged_payments.xlsx with a TOTAL row 'Invoices=..., Payments=...'). The judgement ports, the plumbing does not: the two workbooks become sheet_files/sheet_rows on the sheets server, the matching conventions become SOP-AP-11 on the docs server, and the flagged list plus the two totals go through harness__submit_answer as graded fields. It also fixes what the source's own reference evaluator gets wrong — that evaluator iterates payments, so it flags all 36 legitimately split payments as problems and, being payment-driven, cannot see an invoice with no payment at all or an invoice paid twice in full, while the 17 genuine single-line variances that account for the whole of its own 210.66 total gap go unflagged. Here the exception set is defined at invoice level and is re-derived from the seeded world. Differs from the two existing threeway_match tasks: ppinv-exceptions-mar and tolerance-dialect-mar each match ONE invoice at a time against its PO and receipt and turn on the tolerance rule (SAP unmaintained key = zero vs Oracle blank = unlimited); this one has no PO or receipt in it at all and is a bulk many-to-many reconciliation where the difficulty is match CARDINALITY - one payment covering two invoices, one invoice covering three payments, one payment belonging to neither.", "provenance": "ported", "version": "1.0.0", "walk_len": 8, "walk_servers": ["docs", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 09:34**\n\nHalden Freight's credit controller has been on the phone. They've run their statement of\naccount for Q1 against what they say they've received from us, and they're telling me we're\n**373.50** behind and that they won't confirm the Q2 dedicated-fleet allocation until it's\ncleared. 373.50 on a quarter that size is a rounding error, and I don't believe it — not from\na vendor we paid in instalments twice this quarter.\n\nTheir statement and our payment register extract for the same period are both on the shared\ndrive. Do the proper reconciliation per SOP-AP-11, not their version of it.\n\nI need this two ways round, because two different people action it: anything we still owe them\ngoes into the Q1 accrual on Monday, and anything we've paid twice I want to raise with them\nwhile we still have leverage over the Q2 allocation. Don't hand me one net number.\n", "rubric": {"checks": {"answer_checks": [{"expect": 116438.25, "field": "statement_total_usd", "gt_sql": "SELECT ROUND(SUM(json_extract(cells,'$[3]')),2) FROM sheet_rows WHERE file='halden-statement-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'HF-%'", "tol_abs": 0.05, "type": "number"}, {"expect": 113197.75, "field": "applied_total_usd", "gt_sql": "SELECT ROUND(SUM(json_extract(cells,'$[2]')),2) FROM sheet_rows WHERE file='ap-payment-register-halden-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'PMT-%' AND substr(json_extract(cells,'$[3]'),1,7) IN (SELECT json_extract(cells,'$[0]') FROM sheet_rows WHERE file='halden-statement-2026Q1.xlsx')", "tol_abs": 0.05, "type": "number"}, {"expect": 3240.5, "field": "net_variance_usd", "gt_sql": "SELECT ROUND((SELECT SUM(json_extract(cells,'$[3]')) FROM sheet_rows WHERE file='halden-statement-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'HF-%') - (SELECT SUM(json_extract(cells,'$[2]')) FROM sheet_rows WHERE file='ap-payment-register-halden-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'PMT-%' AND substr(json_extract(cells,'$[3]'),1,7) IN (SELECT json_extract(cells,'$[0]') FROM sheet_rows WHERE file='halden-statement-2026Q1.xlsx')),2)", "tol_abs": 0.05, "type": "number"}, {"expect": ["HF-2203", "HF-2207", "HF-2218", "HF-2222"], "field": "exception_invoices", "forbid": ["HF-2201", "HF-2202", "HF-2204", "HF-2205", "HF-2206", "HF-2208", "HF-2209", "HF-2210", "HF-2211", "HF-2212", "HF-2213", "HF-2214", "HF-2215", "HF-2216", "HF-2217", "HF-2219", "HF-2220", "HF-2221", "HF-2223", "HF-2224", "HF-2190"], "type": "contains_all"}, {"expect": 2750.0, "field": "overpaid_usd", "gt_sql": "WITH stmt AS (SELECT json_extract(cells,'$[0]') AS inv, json_extract(cells,'$[3]') AS amt FROM sheet_rows WHERE file='halden-statement-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'HF-%'), pay AS (SELECT json_extract(cells,'$[3]') AS ref, json_extract(cells,'$[2]') AS amt FROM sheet_rows WHERE file='ap-payment-register-halden-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'PMT-%'), single AS (SELECT substr(ref,1,7) AS inv, SUM(amt) AS paid FROM pay WHERE ref NOT LIKE '%,%' GROUP BY 1), comb AS (SELECT substr(ref,1,7) AS a, substr(ref,9,7) AS b FROM pay WHERE ref LIKE '%,%'), app AS (SELECT s.inv AS inv, s.amt AS amt, COALESCE((SELECT paid FROM single WHERE single.inv=s.inv),0) + CASE WHEN EXISTS(SELECT 1 FROM comb WHERE comb.a=s.inv OR comb.b=s.inv) THEN s.amt ELSE 0 END AS paid FROM stmt s) SELECT ROUND(SUM(paid-amt),2) FROM app WHERE paid > amt", "tol_abs": 0.05, "type": "number"}, {"expect": 5990.5, "field": "underpaid_usd", "gt_sql": "WITH stmt AS (SELECT json_extract(cells,'$[0]') AS inv, json_extract(cells,'$[3]') AS amt FROM sheet_rows WHERE file='halden-statement-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'HF-%'), pay AS (SELECT json_extract(cells,'$[3]') AS ref, json_extract(cells,'$[2]') AS amt FROM sheet_rows WHERE file='ap-payment-register-halden-2026Q1.xlsx' AND json_extract(cells,'$[0]') LIKE 'PMT-%'), single AS (SELECT substr(ref,1,7) AS inv, SUM(amt) AS paid FROM pay WHERE ref NOT LIKE '%,%' GROUP BY 1), comb AS (SELECT substr(ref,1,7) AS a, substr(ref,9,7) AS b FROM pay WHERE ref LIKE '%,%'), app AS (SELECT s.inv AS inv, s.amt AS amt, COALESCE((SELECT paid FROM single WHERE single.inv=s.inv),0) + CASE WHEN EXISTS(SELECT 1 FROM comb WHERE comb.a=s.inv OR comb.b=s.inv) THEN s.amt ELSE 0 END AS paid FROM stmt s) SELECT ROUND(SUM(amt-paid),2) FROM app WHERE paid < amt", "tol_abs": 0.05, "type": "number"}, {"expect": ["PMT-8817", "HF-2190"], "field": "out_of_scope_payment", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"n": 1, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-091-tac-invoice-matching", "task_name": "threeway_match/tac-invoice-matching", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-092-tolerance-dialect-mar/documents/policy--ap-match-tolerances.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"blocked_invoice": "TDINV-401", "blocked_variance": 40.0, "blocking_rule": "legacy ledger: an unmaintained tolerance key means zero tolerance, so any variance blocks", "passing_invoice": "TDINV-402", "passing_variance": 315.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "threeway_match", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-092-tolerance-dialect-mar", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/external/articles: SAP unmaintained tolerance key = zero tolerance vs Oracle Fusion blank tolerance = infinite — inverted defaults across systems", "provenance": "ported", "version": "1.0.0", "walk_len": 5, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 11:02**\n\nTDINV-401 and TDINV-402 both came back with price variances, and both ledgers are telling me no tolerance is configured. Which one can I release and which do I have to hold?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["TDINV-401"], "field": "blocked_invoice", "type": "contains_all"}, {"expect": 40.0, "field": "blocked_variance", "tol_abs": 0.01, "type": "number"}, {"expect": ["TDINV-402"], "field": "passing_invoice", "type": "contains_all"}, {"expect": 315.0, "field": "passing_variance", "tol_abs": 0.01, "type": "number"}, {"expect": ["zero"], "field": "blocking_rule", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-092-tolerance-dialect-mar", "task_name": "threeway_match/tolerance-dialect-mar", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-093-tolerance-dialect-mar-esc-burie-quiet/documents/policy--ap-match-tolerances.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"blocked_invoice": "TDINV-401", "blocked_variance": 40.0, "blocking_rule": "legacy ledger: an unmaintained tolerance key means zero tolerance, so any variance blocks", "passing_invoice": "TDINV-402", "passing_variance": 315.0}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "threeway_match", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-093-tolerance-dialect-mar-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of threeway_match/tolerance-dialect-mar via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: research/external/articles: SAP unmaintained tolerance key = zero tolerance vs Oracle Fusion blank tolerance = infinite — inverted defaults across systems", "provenance": "variant", "version": "1.0.0", "walk_len": 5, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 11:02**\n\nTDINV-401 and TDINV-402 both came back with price variances, and both ledgers are telling me no tolerance is configured. Which one can I release and which do I have to hold?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["TDINV-401"], "field": "blocked_invoice", "type": "contains_all"}, {"expect": 40.0, "field": "blocked_variance", "tol_abs": 0.01, "type": "number"}, {"expect": ["TDINV-402"], "field": "passing_invoice", "type": "contains_all"}, {"expect": 315.0, "field": "passing_variance", "tol_abs": 0.01, "type": "number"}, {"expect": ["zero"], "field": "blocking_rule", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-093-tolerance-dialect-mar-esc-burie-quiet", "task_name": "threeway_match/tolerance-dialect-mar-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-094-bank-change-verify/documents/policy--vendor-bank-change.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"red_flag": "sender domain does not match the vendor contact email on the ERP master record (lookalike domain)", "request_verified": "no", "required_next_step": "call back the vendor on the phone number already on file in the ERP — never a number supplied in the request", "vendor_account": "1001"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-094-bank-change-verify", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "research/finance-agent-workflows.md wf#5 (vendor onboarding/master data + bank-detail fraud checks)", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 08:35**\n\nA vendor's asking us to change their bank details before Friday's run and they're pushing hard on timing. Can we do it? Walk me through what you checked.\n", "rubric": {"checks": {"answer_checks": [{"expect": ["1001"], "field": "vendor_account", "type": "contains_all"}, {"expect": "no", "field": "request_verified", "type": "yes_no"}, {"expect": ["domain"], "field": "red_flag", "type": "contains_all"}, {"expect": ["call"], "field": "required_next_step", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "email", "docs"], "type": "required_servers"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-094-bank-change-verify", "task_name": "vendor_master/bank-change-verify", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-095-dormant-vendor-review/documents/vendor-hygiene-policy.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"cutoff_date": "2025-03-02", "dormant_count": 2, "dormant_vendors": "DVEN-02, DVEN-03", "out_of_scope_vendors": "DVEN-06", "unused_vendors": "DVEN-04"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-095-dormant-vendor-review", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "closes tracked coverage gap art.vendor_dormancy (sim/coverage.py); dormant-account BEC exposure from research/external/articles/vendor-master-bec-fraud--{afp-truist,fbi-ic3,wa-state-auditor}.md; docs/HARD-LAYER-DESIGN.md M3/M5", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Marcus Bell · AP Controls · Teams 09:30**\n\nQuarterly vendor hygiene review for the REVIEW group — can you work out which ones need\ndeactivating? Follow SOP-AP-11 to the letter on this, the definition is fussier than it\nlooks and last quarter's numbers had to be restated because two different findings got\ncounted in the same bucket.\n\nUse today as the review date.\n", "rubric": {"checks": {"answer_checks": [{"expect": 2, "field": "dormant_count", "tol_abs": 0.01, "type": "number"}, {"expect": ["DVEN-02", "DVEN-03"], "field": "dormant_vendors", "forbid": ["DVEN-04", "DVEN-05", "DVEN-06"], "type": "contains_all"}, {"expect": ["2025-03-02"], "field": "cutoff_date", "type": "contains_all"}, {"expect": ["DVEN-04"], "field": "unused_vendors", "type": "contains_all"}, {"expect": ["DVEN-06"], "field": "out_of_scope_vendors", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-095-dormant-vendor-review", "task_name": "vendor_master/dormant-vendor-review", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-096-dormant-vendor-review-esc-burie-quiet/documents/vendor-hygiene-policy.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"cutoff_date": "2025-03-02", "dormant_count": 2, "dormant_vendors": "DVEN-02, DVEN-03", "out_of_scope_vendors": "DVEN-06", "unused_vendors": "DVEN-04"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-096-dormant-vendor-review-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of vendor_master/dormant-vendor-review via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: closes tracked coverage gap art.vendor_dormancy (sim/coverage.py); dormant-account BEC exposure from research/external/articles/vendor-master-bec-fraud--{afp-truist,fbi-ic3,wa-state-auditor}.md; docs/HARD-LAYER-DESIGN.md M3/M5", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Marcus Bell · AP Controls · Teams 09:30**\n\nQuarterly vendor hygiene review for the REVIEW group — can you work out which ones need\ndeactivating? Follow the governing policy to the letter on this, the definition is fussier than it\nlooks and last quarter's numbers had to be restated because two different findings got\ncounted in the same bucket.\n\nUse today as the review date.\n", "rubric": {"checks": {"answer_checks": [{"expect": 2, "field": "dormant_count", "tol_abs": 0.01, "type": "number"}, {"expect": ["DVEN-02", "DVEN-03"], "field": "dormant_vendors", "forbid": ["DVEN-04", "DVEN-05", "DVEN-06"], "type": "contains_all"}, {"expect": ["2025-03-02"], "field": "cutoff_date", "type": "contains_all"}, {"expect": ["DVEN-04"], "field": "unused_vendors", "type": "contains_all"}, {"expect": ["DVEN-06"], "field": "out_of_scope_vendors", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-096-dormant-vendor-review-esc-burie-quiet", "task_name": "vendor_master/dormant-vendor-review-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-097-dormant-vendor-review-v2/documents/vendor-hygiene-policy.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"cutoff_date": "2025-03-02", "dormant_count": 2, "dormant_vendors": "DVEN-02, DVEN-03", "out_of_scope_vendors": "DVEN-06", "unused_vendors": "DVEN-04"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-097-dormant-vendor-review-v2", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of vendor_master/dormant-vendor-review via the tau2.retrieval_modes lever (doc_mode=buried, sim/prepare.py): identical ground truth, harder retrieval. Base task closes art.vendor_dormancy (sim/coverage.py); dormant-account BEC exposure from research/external/articles/vendor-master-bec-fraud--{afp-truist,fbi-ic3,wa-state-auditor}.md; docs/HARD-LAYER-DESIGN.md M3/M5", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Marcus Bell · AP Controls · Teams 09:30**\n\nQuarterly vendor hygiene review for the REVIEW group — can you work out which ones need\ndeactivating? Whatever the current hygiene policy says, follow it to the letter - the definition is\nfussier than it looks and last quarter's numbers had to be restated because two different\nfindings got counted in the same bucket. I can never remember which SOP number it is, there\nare a lot of them on that shelf now.\n\nUse today as the review date.\n", "rubric": {"checks": {"answer_checks": [{"expect": 2, "field": "dormant_count", "tol_abs": 0.01, "type": "number"}, {"expect": ["DVEN-02", "DVEN-03"], "field": "dormant_vendors", "forbid": ["DVEN-04", "DVEN-05", "DVEN-06"], "type": "contains_all"}, {"expect": ["2025-03-02"], "field": "cutoff_date", "type": "contains_all"}, {"expect": ["DVEN-04"], "field": "unused_vendors", "type": "contains_all"}, {"expect": ["DVEN-06"], "field": "out_of_scope_vendors", "type": "contains_all"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["erp", "docs"], "type": "required_servers"}, {"n": 2, "server": "erp", "type": "min_calls"}, {"type": "reads_before_submit"}, {"n": 3, "server": "docs", "type": "min_calls"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-097-dormant-vendor-review-v2", "task_name": "vendor_master/dormant-vendor-review-v2", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 1, "name": "agent_sent_mail", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent'", "type": "row_count"}, {"expect": 1, "name": "vendor_replied", "sql": "SELECT COUNT(*) FROM email_messages WHERE id='em-reply-wingtip-po'", "type": "row_count"}, {"expect": 1, "name": "mailed_the_vendor_on_file", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent' AND to_addr='ap@wingtip-sim.example'", "type": "row_count"}], "submit_answer": {"contact_emailed": "ap@wingtip-sim.example", "po_exists_in_erp": "yes", "po_item": "PKG-LABEL-ROLL", "po_number": "PO-7003"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-098-missing-po-inquiry", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "TheAgentCompany finance suite pattern: agent must message an NPC for a blocking fact (research/external/repos/INDEX.md); enabled by email.send_message + scripted counterparties", "provenance": "ported", "version": "1.0.0", "walk_len": 6, "walk_servers": ["email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 09:22**\n\nTWINV-303 from Wingtip has no PO reference on it so it won't match or schedule. Can you sort it out and confirm the PO is real on our side?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PO-7003"], "field": "po_number", "type": "contains_all"}, {"expect": "yes", "field": "po_exists_in_erp", "type": "yes_no"}, {"expect": ["PKG-LABEL-ROLL"], "field": "po_item", "type": "contains_all"}, {"expect": ["wingtip"], "field": "contact_emailed", "type": "contains_all"}], "state_checks": [{"tables": ["answers", "email_messages"], "type": "writes_only"}, {"expect": 1, "name": "agent_sent_mail", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent'", "type": "row_count"}, {"expect": 1, "name": "vendor_replied", "sql": "SELECT COUNT(*) FROM email_messages WHERE id='em-reply-wingtip-po'", "type": "row_count"}, {"expect": 1, "name": "mailed_the_vendor_on_file", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent' AND to_addr='ap@wingtip-sim.example'", "type": "row_count"}], "trace_checks": [{"servers": ["erp", "email"], "type": "required_servers"}, {"n": 2, "server": "email", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-098-missing-po-inquiry", "task_name": "vendor_master/missing-po-inquiry", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": [], "gold_output": {"expected_state_assertions": [{"expect": 1, "name": "agent_sent_mail", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent'", "type": "row_count"}, {"expect": 1, "name": "vendor_replied", "sql": "SELECT COUNT(*) FROM email_messages WHERE id='em-reply-wingtip-po'", "type": "row_count"}, {"expect": 1, "name": "mailed_the_vendor_on_file", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent' AND to_addr='ap@wingtip-sim.example'", "type": "row_count"}], "submit_answer": {"contact_emailed": "ap@wingtip-sim.example", "po_exists_in_erp": "yes", "po_item": "PKG-LABEL-ROLL", "po_number": "PO-7003"}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-099-missing-po-inquiry-esc-burie-quiet", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "escalated variant of vendor_master/missing-po-inquiry via sim/grow_tasks.py (buried_docs: governing policy hidden in the adjacent-policy library; quiet_prompt: explicit policy pointers removed from the ask); ground truth unchanged. Base: TheAgentCompany finance suite pattern: agent must message an NPC for a blocking fact (research/external/repos/INDEX.md); enabled by email.send_message + scripted counterparties", "provenance": "variant", "version": "1.0.0", "walk_len": 6, "walk_servers": ["email", "erp", "harness"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Priya Shah · AP Manager · Teams 09:22**\n\nTWINV-303 from Wingtip has no PO reference on it so it won't match or schedule. Can you sort it out and confirm the PO is real on our side?\n", "rubric": {"checks": {"answer_checks": [{"expect": ["PO-7003"], "field": "po_number", "type": "contains_all"}, {"expect": "yes", "field": "po_exists_in_erp", "type": "yes_no"}, {"expect": ["PKG-LABEL-ROLL"], "field": "po_item", "type": "contains_all"}, {"expect": ["wingtip"], "field": "contact_emailed", "type": "contains_all"}], "state_checks": [{"tables": ["answers", "email_messages"], "type": "writes_only"}, {"expect": 1, "name": "agent_sent_mail", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent'", "type": "row_count"}, {"expect": 1, "name": "vendor_replied", "sql": "SELECT COUNT(*) FROM email_messages WHERE id='em-reply-wingtip-po'", "type": "row_count"}, {"expect": 1, "name": "mailed_the_vendor_on_file", "sql": "SELECT COUNT(*) FROM email_messages WHERE folder='sent' AND to_addr='ap@wingtip-sim.example'", "type": "row_count"}], "trace_checks": [{"servers": ["erp", "email"], "type": "required_servers"}, {"n": 2, "server": "email", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-099-missing-po-inquiry-esc-burie-quiet", "task_name": "vendor_master/missing-po-inquiry-esc-burie-quiet", "world_id": "ledgerbench-erp-world-v1"}
+{"context_files": ["task_files/lgr100-100-tac-find-signatories/documents/memo--interim-signing-delegation.md", "task_files/lgr100-100-tac-find-signatories/documents/policy--treasury-signing-authority.md"], "gold_output": {"expected_state_assertions": [], "submit_answer": {"authorised_signatories": "Marguerite Okonjo, Halvard Brekke, Priya Shah, Desmond Achebe", "authorised_signatory_count": 4, "departed_signatory": "SIG-06, Corinne Vasquez, Director of Financial Reporting (Band B, USD 250,000). The HR extract records her as a leaver with a final working day of 2026-02-13, so her authority fell away on that date under SOP-TR-02 s2.1; the register was never updated and her offboarding checklist is still open. Strike the row and revoke her payment-release access.", "expired_delegation": "SIG-07, held by Toma Larsen (Band B, USD 250,000, interim cover for the Controller). The CFO's delegation of 2025-11-03 was time-limited and ended 2026-02-28, so the authority is void from 2026-03-01 under SOP-TR-02 s2 even though the register still shows the row Active. He remains an employee in his substantive role; only the signing authority has lapsed, and the row must be struck at this refresh.", "max_release_without_band_a": 50000, "valid_mandate_rows": 5}}, "metadata": {"benchmark": "LedgerBench-100", "code_license": "Apache-2.0", "data_license": "CC-BY-4.0", "difficulty": "hard", "family": "vendor_master", "grading": "deterministic", "harbor_name": "blobfishai/lgr100-100-tac-find-signatories", "llm_judge": false, "mcp_servers": 8, "mcp_tools": 66, "origin": "port of research/external/repos/TheAgentCompany/workspaces/tasks/finance-find-signatories (establish the set of signatories from source records, then report both the number of signatures found and the smaller number of PEOPLE to act on - the original's 38 signatures vs 29 signatories). The original's ownCloud/RocketChat plumbing does not port: the workbooks become sheets (bank mandate register + HR personnel extract), the rules become docs (SOP-TR-02 and the CFO's interim delegation memo), and the answer goes through harness__submit_answer. Mechanics per docs/HARD-LAYER-DESIGN.md: M3 non-collapse (the register's own Active column is the naive source of truth and is wrong in two directions), M5 the validity rules live only in the SOP and the delegation memo - no field on the register says an authority has fallen away, and the Register status column actively asserts the opposite.", "provenance": "ported", "version": "1.0.0", "walk_len": 9, "walk_servers": ["docs", "email", "harness", "sheets"], "world_epoch": "2026-03-02T12:00:00Z"}, "prompt": "**Nyla Ferreira · Treasury Manager · Teams 08:20**\n\nFirst National want the mandate on the operating account re-certified by Friday, and anyone who\nisn't on the list I send gets struck off and their release access closed the same day. So I'd\nrather not just forward them the register as it stands.\n\nTreasury Ops keep that register by hand and I don't trust the status column on it — there's\nbeen movement in Reporting this quarter, and Halvard is back from leave, so the cover\narrangements we put in over the autumn have unwound. HR's finance-function extract is on the\nshared drive next to it.\n\nWork it per SOP-TR-02 and tell me who can genuinely sign today. I also need whatever has to come\noff the mandate, by name, so I can raise the ERP access revocations in the same pass.\n\nOne more for Friday's run: there's a supplier payment that has been going out on a Band B\nsignature. Tell me the largest single payment we can still release without going to the CFO or\nto Halvard.\n", "rubric": {"checks": {"answer_checks": [{"expect": 5, "field": "valid_mandate_rows", "gt_sql": "SELECT COUNT(*) FROM sheet_rows m JOIN sheet_rows h ON json_extract(h.cells,'$[0]')=json_extract(m.cells,'$[1]') WHERE m.file='bank-mandate-register-2026.xlsx' AND h.file='hr-personnel-extract-2026-03-01.xlsx' AND json_extract(m.cells,'$[0]') LIKE 'SIG-%' AND json_extract(h.cells,'$[4]')='Active' AND json_extract(m.cells,'$[6]')<=substr((SELECT value FROM meta WHERE key='WORLD_NOW'),1,10) AND (json_extract(m.cells,'$[7]')='' OR json_extract(m.cells,'$[7]')>=substr((SELECT value FROM meta WHERE key='WORLD_NOW'),1,10))", "tol_abs": 0.01, "type": "number"}, {"expect": 4, "field": "authorised_signatory_count", "gt_sql": "SELECT COUNT(DISTINCT json_extract(m.cells,'$[1]')) FROM sheet_rows m JOIN sheet_rows h ON json_extract(h.cells,'$[0]')=json_extract(m.cells,'$[1]') WHERE m.file='bank-mandate-register-2026.xlsx' AND h.file='hr-personnel-extract-2026-03-01.xlsx' AND json_extract(m.cells,'$[0]') LIKE 'SIG-%' AND json_extract(h.cells,'$[4]')='Active' AND json_extract(m.cells,'$[6]')<=substr((SELECT value FROM meta WHERE key='WORLD_NOW'),1,10) AND (json_extract(m.cells,'$[7]')='' OR json_extract(m.cells,'$[7]')>=substr((SELECT value FROM meta WHERE key='WORLD_NOW'),1,10))", "tol_abs": 0.01, "type": "number"}, {"expect": ["Okonjo", "Brekke", "Shah", "Achebe"], "field": "authorised_signatories", "forbid": ["Vasquez", "Larsen", "Fabbri", "Nystrom", "Bergstrom", "Ferreira"], "type": "contains_all"}, {"expect": ["Larsen"], "field": "expired_delegation", "forbid": ["Vasquez"], "type": "contains_all"}, {"expect": ["Vasquez"], "field": "departed_signatory", "forbid": ["Larsen", "Bergstrom"], "type": "contains_all"}, {"expect": 50000.0, "field": "max_release_without_band_a", "gt_sql": "SELECT MAX(json_extract(m.cells,'$[5]')) FROM sheet_rows m JOIN sheet_rows h ON json_extract(h.cells,'$[0]')=json_extract(m.cells,'$[1]') WHERE m.file='bank-mandate-register-2026.xlsx' AND h.file='hr-personnel-extract-2026-03-01.xlsx' AND json_extract(m.cells,'$[0]') LIKE 'SIG-%' AND json_extract(m.cells,'$[4]')<>'A' AND json_extract(h.cells,'$[4]')='Active' AND json_extract(m.cells,'$[6]')<=substr((SELECT value FROM meta WHERE key='WORLD_NOW'),1,10) AND (json_extract(m.cells,'$[7]')='' OR json_extract(m.cells,'$[7]')>=substr((SELECT value FROM meta WHERE key='WORLD_NOW'),1,10))", "tol_abs": 0.05, "type": "number"}], "state_checks": [{"tables": ["answers"], "type": "writes_only"}], "trace_checks": [{"servers": ["sheets", "docs"], "type": "required_servers"}, {"n": 3, "server": "sheets", "type": "min_calls"}, {"n": 2, "server": "docs", "type": "min_calls"}, {"type": "reads_before_submit"}]}, "engine": "verifiers/vcode.py (binary reward; all checks must pass)", "gates": ["answer_checks: submitted fields graded by type with tolerances", "trace_checks: required servers visited, reads precede submission", "state_checks: writes_only anti-hack veto plus SQL over the world left behind"], "type": "deterministic"}, "task_id": "lgr100-100-tac-find-signatories", "task_name": "vendor_master/tac-find-signatories", "world_id": "ledgerbench-erp-world-v1"}