qa/ctx_benchmark_status.csv

id,area,user_story,expected_behavior,status,evidence,repro,risk,fix,reviewer_verdict,last_updated
"BENCH-001","repo-cache","As a benchmark operator I can prepare pinned public repositories offline after one fetch.","The exact Click and Requests commits have complete source blobs and every arm clones the same commit.","Resolved","Commit 25a3c082; red and reference controls passed in .gate/ctx-ab-runs/ctx-ab-20260727T221130Z-7af605ce and local-fast passed on committed HEAD.","Run the private calibration scenario with retries 0 and inspect controls/control.json before either arm.","Partial caches can fail later when a blob is first read.","Use a full mirror and verify or fetch the exact commit before preparing an arm.","Pass: pinned repositories, evaluator controls, and committed-head package gates are green.","2026-07-28"
"BENCH-002","path-isolation","As an operator I can benchmark without writing ctx state into my normal home directory.","Each run receives an isolated owner-only HOME and temporary directory; authentication is copied with mode 0600 only for the turn and the entire agent home is removed in finally.","Resolved","Both arms in .gate/ctx-ab-runs/ctx-ab-20260727T221130Z-7af605ce passed evaluator-isolation.json: network denied, git and project canaries passed, and credential, config, scenario, control, and sibling paths were denied.","Run one private production pair, inspect evaluator-isolation.json for each arm, and verify no auth.json or config.toml remains under the run root.","Persistent credentials or shared user paths can contaminate evidence and expose secrets.","Use a 0700 per-arm home, 0600 credential/config copies, explicit deny rules, and fail-closed cleanup.","Pass: independent security review accepted the isolation boundary and artifact scan.","2026-07-28"
BENCH-003,recommendations,"As a reviewer I can distinguish recommendation selection and use.","All candidates are persisted; only policy-selected entities load and only trace-proven entities become used.",Resolved,"Dry-run summary recommends skill agent and MCP but ctx-light selects only its skill; lifecycle contains only that skill.","Inspect click ctx-light recommendations.json summary.json and lifecycle/events.jsonl in the adaptive dry run.","Treating every recommendation as selected creates fake lifecycle evidence and needless cost.","Persist recommended_ids selected_ids and used_ids separately and drive lifecycle from selected_items.","Pass after fix: candidate and active-context states no longer conflict.",2026-07-20
"BENCH-004","operational-context","As a reviewer I can prove which skills MCPs and agents were loaded used and unloaded.","Light mode loads one skill; full mode loads all three; MCP and agent use require exact runtime evidence; only loaded entities unload.","Resolved","Production dry runs .gate/ctx-ab-runs/ctx-ab-20260727T223036Z-1d7e43ad and ctx-ab-20260727T223114Z-0be5b69e record candidates, selection, lifecycle, final empty state, and cache provenance.","Run a dry run and inspect recommendations.json, summary.json, and lifecycle actions for each arm.","Lifecycle-only claims can misrepresent tools the host never supplied or used.","Require selected, delivered, adopted, and used states separately; unload only applied entities.","Pass for protocol and production-catalog wiring; live full runtime proof remains BENCH-011.","2026-07-28"
BENCH-005,token-accounting,As a product owner I receive honest time and token KPIs.,"Parent and orchestration totals preserve provider-reported input, cache-read, cache-write, uncached, and output tokens; missing provider usage remains unavailable and completeness is explicit.",Resolved,"Commits 43fda102, 0708979d, 0854ac70, and fc4cf0b9 preserve cumulative usage, cache-read/cache-write detail, history, and failure-as-unavailable behavior; the orchestration checkpoint passed 5140 tests.","Exercise complete, partial, timeout, exception, cache-read, and cache-write provider records, then compare CLI, session history, lifecycle, telemetry, and dashboard totals.",Zero-filling missing usage or dropping cache-write tokens understates cost and invalidates enterprise KPI comparisons.,"Use complete-or-null aggregation, persist tokens_reported and cache fields, and keep unavailable provider failures null across every surface.",Pass for truthful shared usage accounting; production benchmark validity remains tracked separately.,2026-07-25
"BENCH-006","fairness","As a product owner I can compare without-ctx and adaptive ctx behavior on the same task.","Arms share commit, model, task, evaluator, timeout, isolation, and verification; six paired trials alternate baseline-light order three times each and exclude incomplete pairs from efficiency claims.","Resolved","Commit 25a3c082 adds explicit two-arm 3/3 counterbalancing, clean-worktree enforcement, immutable scenario and catalog hashes, and paired completeness tests; Click runs are labeled calibration only.","Run six trials from a clean committed harness and inspect the schedule, immutable hashes, performance.json, summary.json, and incidents.csv.","Dirty harnesses, fixed arm order, or asymmetric failures can create a false efficiency result.","Fail closed on dirty state, alternate order, preserve intent-to-treat quality failures, and calculate efficiency only for complete passing pairs.","Pass for benchmark methodology; production benefit evidence remains open in BENCH-008, BENCH-014, and BENCH-021.","2026-07-28"
BENCH-007,retry-loop,"As an operator I get a durable record of failures and bounded retries.","Failures are logged and later correlated retries resolve them; unresolved incidents fail; light escalates only after a clean model turn and normal verifier exit 1.",Needs Validation,"Unit tests cover correlated resolution and timeout/non-test failure non-escalation; adaptive live recovery evidence is pending.","Inject a normal focused-test failure and a timeout then inspect treatment levels and incident statuses.","Silent retries hide instability and timeout escalation activates unnecessary expensive context.","Resolve prior attempt rows after success reject unresolved rows and gate escalation on exact failure kind.","Review blockers fixed in code; live recovery remains.",2026-07-20
"BENCH-008","treatment-overhead","As a ctx user I receive useful context without paying for every available tool.","Host preflight selects at most one bounded skill or abstains; expensive MCP and reviewer surfaces activate only after explicit evidence; selected context is removed after its useful phase.","Needs Validation","Historical Click calibration favored CTX, while the six-pair attrs diagnostic preserved quality but did not meet the uncached-token benefit threshold. Those exposed, outcome-informed assignments are diagnostic only. The official V2 campaign remains NOT RUN: 0/30 pairs and 0/60 arms.","In the frozen official V2 campaign, inspect every CTX arm for exact recommendation, selection, load, use, unload, abstention, and provider-token evidence, then apply the preregistered repository-level quality and benefit gates.","Generic guidance or unnecessary tool surfaces can preserve quality while increasing uncached tokens and elapsed time.","Keep bounded trace metrics, fail closed on run attestation, load only policy-selected context, and require quality non-inferiority plus the preregistered repository-level benefit threshold.","Keep open until the official task-disjoint V2 campaign reports all 30 pairs; do not reuse the exposed attrs, Requests, or Rich assignments as confirmatory evidence.","2026-07-30"
"BENCH-009","arm-order","As a reviewer I get order-balanced results even when scenarios are filtered.","Each six-trial three-arm block uses every permutation once; each six-trial baseline-light block alternates order exactly three times each.","Resolved","Commit 25a3c082; test_three_arm_schedule_is_stable_and_counterbalanced and test_two_arm_schedule_alternates_order_three_times_each pass.","Compare full and filtered three-arm schedules, then assert the two-arm six-trial sequence is baseline-first and CTX-first 3/3.","Resetting scenario indexes or always starting one arm biases timing and cache effects.","Use stable scenario ordering and explicit two-arm alternation.","Pass: both supported schedules are deterministic and counterbalanced.","2026-07-28"
BENCH-010,process-containment,"As an operator I do not retain detached benchmark or verification processes.","Timeout and successful-parent paths terminate descendants on supported POSIX hosts including detached children that retain inherited pipes and fail closed when cleanup cannot be verified.",Needs Validation,"HISTORICAL PLATFORM FACT: code commit ea3625f6 added a kill-on-close Windows Job Object, suspended root assignment before resume, dynamic PID verification, cleanup-before-pipe-drain behavior, verified emergency cleanup, and native success and timeout regressions. The exact combined V2 suite passed 505/505 on code HEAD 547b80a5; independent runtime and source-boundary reviews approved that candidate. CURRENT POSIX CANDIDATE: native Windows containment was removed and process-group plus marker cleanup remains the supported contract.","Run the POSIX detached-descendant successful-parent, inherited-pipe, timeout, and cleanup-failure regressions, then scan every recorded PID after the official campaign.","Detached children distort elapsed time, leak resources, or turn a successful root into a false timeout while retaining inherited pipes.","Use POSIX process-group plus marker cleanup; terminate and verify containment on every success, timeout, and exception path before draining inherited pipes.","Supported-host implementation and focused gates pass; the official post-run PID scan remains required.",2026-08-01
BENCH-011,runtime-tool-proof,"As a reviewer I can reject an MCP or reviewer claim that used the wrong target or was not closed.","Full treatment passes only with the expected MCP slug type body and a matching selected-marker spawned waited and closed reviewer ID.",Needs Validation,"Unit fixtures reject failed wrong-argument wrong-result unclosed and malicious-review prompts; real MCP protocol preflight passes.","Run one live ctx-full arm and inspect policy_valid tool observations marker and team-token scope.","Tool-name-only evidence can credit the wrong call and completed reviewers can remain allocated.","Bind MCP evidence to exact arguments/result bind agent evidence to selected marker and mark any agent attempt token-incomplete.","Review blocker fixed in code; final live trace pending.",2026-07-20
"BENCH-012","run-manifest","As a reviewer I can reproduce the exact benchmark schedule and repository state.","The manifest records trials, retries, timeout, arm mode, filters, cache, schedule, script and scenario hashes, dependencies, catalog provenance, and clean state; live scoring rejects dirty state.","Resolved","Commit 25a3c082 and .gate/ctx-ab-runs/ctx-ab-20260727T223114Z-0be5b69e/environment.json record the version-2 cache hit, seven bound runtime files, schedule, hashes, and dirty diagnostic state.","Run a dry run from a dirty tree, inspect environment.json, then confirm the same live command fails before model execution.","Missing parameters or allowing a changing harness makes evidence irreproducible.","Persist the complete manifest and require a clean committed harness for live production runs.","Pass: manifest completeness and dirty-tree fail-closed behavior are independently reviewed.","2026-07-28"
BENCH-013,feature-registry,"As a ctx maintainer I can discover the benchmark from the repository feature inventory.","The benchmark runner has a user story and evidence row in both canonical and published feature trackers.",Resolved,"MAINT-016 exists in qa/feature_status.csv and docs/qa/feature-user-story-status.csv; all 17 feature tracker tests pass.","Run .venv/bin/python -m pytest -q src/tests/test_feature_user_story_tracker.py.","An untracked maintainer surface violates the repository-wide feature audit goal and fails CI.","Add one MAINT-016 row to each approved tracker and retain the benchmark ledger as source evidence.","Pass: canonical and published rows parse with unique IDs and the maintainer-script coverage gate is green.",2026-07-30
"BENCH-014","catalog-validity","As a product owner I can distinguish controlled context delivery from real-catalog recommendation quality.","Controlled fixtures prove mechanics; production runs use the shipped graph, availability filtering, exact source-body provenance, and a separate held-out relevance evaluation.","Needs Validation","Commit 25a3c082 binds all seven declared runtime files byte-for-byte to the hashed availability pack on build and cache hit. Dry runs ctx-ab-20260727T223036Z-1d7e43ad and ctx-ab-20260727T223114Z-0be5b69e prove cache miss/hit and selected installable context. A frozen held-out relevance set is still absent.","Freeze private scenarios before CTX inspection, run red/reference controls, then evaluate relevance, availability, source safety, language fit, no-key constraints, and paired outcomes.","Mechanically valid catalog entries can still be irrelevant or overly lexical, and calibration tasks cannot serve as holdouts.","Retain exact source binding and add private held-out assignments across at least three repositories before any broad quality claim.","Keep open: catalog mechanics and provenance pass; held-out multi-repository recommendation and development evidence remains mandatory.","2026-07-28"
BENCH-015,end-to-end-time,"As a product owner I can compare true elapsed development time rather than selected phases.","Total time includes clone preparation ctx setup model work verification artifact checks lifecycle unload and session close while phase times remain visible.",Resolved,"Runner now measures from trial start through lifecycle close and materialization elapsed is included in verification; focused tests and dry run pass.","Compare total_seconds with phase fields and teardown_seconds in an adaptive run.","Excluding ctx teardown or artifact work understates treatment cost.","Measure end-to-end with a monotonic timer and retain setup agent verification and teardown breakdowns.","Reviewer timing gap fixed in code; official live values remain open under BENCH-035.",2026-07-20
BENCH-016,reviewer-order-proof,"As a benchmark reviewer I can trust that a reviewer agent really completed its assigned review.","Full treatment credits reviewer use only when an error-free selected spawn is followed by a completed non-empty wait and close for the same agent in that order.",Resolved,"Adversarial fixtures reject completed calls carrying errors and close-wait-spawn ordering; 21 focused tests pass.","Feed observed_agent_review a completed event with a non-null error or reverse the valid spawn-wait-close events.","Set intersection without ordering can falsely certify failed or impossible reviewer traces.","Track each selected reviewer through an ordered spawn-wait-close state machine and ignore every error-bearing event.","Pass: independent recheck closed the P1 with valid error-stage mismatched-ID reverse-order and success probes.",2026-07-20
BENCH-017,runtime-surface-isolation,"As a ctx user I do not pay prompt or tool overhead for capabilities that were not selected.","Baseline and ctx-light explicitly disable multi-agent and expose no ctx MCP configuration; only ctx-full enables both surfaces.",Resolved,"Command-construction regression checks disable/no-MCP for light and enable/MCP for full; focused gate passes.","Build Codex commands for with_ctx false and true then inspect feature flags and MCP config arguments.","Merely exposing unused agent or MCP schemas can inflate cached context and lets baseline behavior escape the intended policy.","Explicitly disable multi_agent outside full treatment and attach the MCP only to full treatment.","Pass: independent product-path probes confirmed recommendation does not auto-load entities and the old harness forced the expensive surfaces.",2026-07-20
BENCH-018,activation-policy,"As a host integrator I receive a bounded adaptive activation proposal instead of an instruction to load every candidate.","context_policy proposes at most one local directly usable skill for immediate load; active baseline external malformed and expensive candidates stay kept manual or deferred.",Resolved,"Commit 98b3ee7f; 79 adapter tests and the 435-test recommendation/CLI integration slice passed.","Call ctx__recommend_bundle with active baseline external malformed skill agent and MCP rows then inspect keep load deferred and manual.","A naive host can interpret the advisory load list literally and recreate the forced-full slowdown in production.","Load one eligible skill; exclude kept context; defer agents MCPs and secondary skills; fail closed on external or malformed metadata.","Approved after two review rounds and direct adversarial probes.",2026-07-20
BENCH-019,tool-schema-budget,"As a ctx run user I send only the ctx schemas configured for the current stage of work.","Fresh runs default to recommend_bundle plus wiki_get; explicit full or allow patterns expose more; prompts and executors match submitted schemas; legacy and overridden resume settings remain durable.",Resolved,"Commit 743d3f20; default schema payload fell from 13 tools and 12476 bytes to 2 tools and 3660 bytes (70.7 percent); 143 CLI/state and 435 integration tests passed.","Run default full no-ctx MCP-only lifecycle-only legacy resume and full-to-minimal resume probes; compare submitted names prompts execution and session_config.","Repeated unused schemas consume input and cached-context tokens on every model iteration.","Filter schemas before provider submission rebuild instructions from actual definitions preserve legacy full resumes and persist explicit resume overrides.","Approved after three review rounds including exact pre-patch prompt migration.",2026-07-20
BENCH-020,legacy-skill-router,"As a developer I receive a small ranked skill selection rather than a bulk context merge.","The legacy CLI defaults to one installed directly mapped skill; fuzzy and graph entities remain suggestions; always_load and explicit broader caps are deterministic and documented.",Resolved,"Commit bcd2e3a6; 51 resolver tests plus 435 integration tests passed; 20 hash-seed probes selected the same conflict winner.","Run C-language fuzzy unavailable-Docker graph-agent Jest-Vitest conflict invalid-limit and explicit-cap probes.","Bulk or lexical skill injection increases prompt size and can mix conflicting instructions.","Use token-boundary fuzzy matching direct-only single-slot activation deterministic ties and named-selection documentation.","Approved after reviewer reproduced and closed all four findings.",2026-07-20
"BENCH-021","adaptive-runtime-context","As a ctx run user I have selected skills, MCPs, and agents activated and unloaded as task evidence changes without exposing every capability up front.","The host control plane recommends outside model turns; each provider turn receives bounded selected context; lifecycle distinguishes selection, delivery, adoption, use, and unload; final state is empty.","Needs Validation","Commit 70dd4917 fixed delivery and terminal attestation. Commit d4786ccb added a pure default-deny deferred decision policy with strict public-evidence schemas, explicit safety flags, provenance, permissions, risk, and recursive oracle rejection; 116 benchmark tests and local-fast passed 10 lanes with 5737 tests passing.","Run held-out private production pairs and require exact lifecycle ordering, cleanup, final empty state, and naturally policy-triggered MCP or reviewer execution in a separate conformance suite.","Prompt delivery alone does not prove adoption or use; the decision policy is implemented but the performance engine does not yet execute an approved deferred MCP or reviewer.","Execute only policy-approved deferred context through host-owned pre-solve MCP and post-solve reviewer phases, keep lifecycle conformance separate from performance claims, and never inject hidden oracle data or forced IDs.","Keep open: skill delivery, cleanup, and deferred decisions are fixed; deferred execution plus live conformance evidence remain required.","2026-07-28"
"BENCH-022","provenance-attestation","As a benchmark reviewer I can prove which provider, model, ctx runtime, source revision, catalog source, and treatment produced every scored result.","Completed trials and aggregates fail closed unless provider, runtime, harness, scenario, catalog, availability pack, generated runtime files, and treatment provenance are internally consistent.","Resolved","Commits a45b6034 and 25a3c082; 90 focused tests cover missing, tampered, mismatched, dirty, and valid provenance. The adversarial installer/source mismatch is rejected and the reviewer returned APPROVE.","Alter provider or runtime attestation, generated runtime file bytes, availability JSON, harness state, or arm treatment and validate the run.","Unattested or falsely bound evidence can measure different code or context while appearing reproducible.","Bind every declared runtime file to source bytes, version cache keys, require clean committed live runs, and exclude unverifiable trials.","Pass: independent adversarial re-review accepted build and cache-hit provenance; efficiency collection remains separately open.","2026-07-28"
"BENCH-023","evaluator-isolation","As a benchmark reviewer I can trust that the model cannot inspect evaluator controls, credentials, sibling prompts, or unrestricted network.","Each arm proves exact sandbox denies while allowing git and the untouched project regression command; no sensitive artifact survives cleanup.","Resolved","Both evaluator-isolation.json files in ctx-ab-20260727T221130Z-7af605ce report verified=true, network canary denied, git=0, project=0, and every sensitive read denied.","Run one production pair and inspect per-arm evaluator-isolation.json plus a post-run sensitive filename scan.","Readable controls leak answers; copied credentials or sibling prompts break security and causal isolation.","Use per-arm 0700 homes, 0600 transient auth, explicit deny paths, loopback canaries, project canaries, and finally cleanup.","Pass: independent security and benchmark reviewers accepted the isolation evidence.","2026-07-28"
"BENCH-024","clean-harness-gate","As a benchmark operator I cannot accidentally publish evidence from changing code.","Live production execution starts only from a clean committed worktree and records immutable harness and scenario hashes.","Resolved","Commit 25a3c082 rejected all five dirty paths before model execution; local-fast passed all 10 lanes with 9 workers in 41.194 seconds; clean canary ctx-ab-20260727T224009Z-61640368 records repository clean=true and an empty tracked diff hash.","Modify one tracked file and run the private live command; it must exit before catalog/model work. Commit, rerun local-fast, and verify all lane return codes are zero before the canary.","Dirty evidence cannot be reproduced or attributed to the merged implementation.","Fail closed on dirty repository state and run all local gates from isolated committed-HEAD worktrees.","Pass: reviewer accepted the gate and the first frozen clean canary exercised it successfully.","2026-07-28"
"BENCH-025","hidden-oracle-governance","As a benchmark owner I can keep real evaluator tests and reference patches unavailable to agents and public source.","Historical or held-out scenarios live only in an ignored 0700 directory with 0600 files; tracked scenario packs contain public synthetic examples only.","Resolved","The Click historical scenario is absent from tracked benchmarks/ctx_ab/scenarios.yaml and preserved at .gate/ctx-ab-private/scenarios.yaml with directory mode 0700 and file mode 0600; live path validation rejects public, temporary, symlinked, or broad-permission sources.","Search tracked files for the historical scenario ID, inspect private modes, and attempt a live run with the public scenario pack.","Public hidden tests invalidate holdouts and readable local files leak evaluator answers.","Separate calibration from confirmation, enforce private roots and owner-only modes, and deny scenario/control paths in the agent sandbox.","Pass: CTO review closed the tracked-oracle blocker.","2026-07-28"
"BENCH-026","trace-efficiency-evidence","As a product owner I can explain token overhead using measured prompt and tool behavior rather than guesses.","Every live arm records prompt bytes, command count, failures, total and maximum tool output, oversized output count, repeated command count, messages, and exact terminal token fields.","Resolved","Commit 25a3c082 and clean calibration ctx-ab-20260727T224009Z-61640368: CTX emitted 37628 tool-output bytes with a 7876-byte maximum and zero oversized commands; baseline emitted 75802 bytes with one 48718-byte command.","Inject repeated and greater-than-32KiB command outputs into JSONL, then inspect summary metrics and the real calibration trace.","Aggregate token counts alone cannot identify context bloat or distinguish prompt overhead from tool-output amplification.","Persist bounded trace metrics and predeclare a 32KiB per-command ceiling for confirmatory trials.","Pass: forensic reviewer accepted the measurements while correctly rejecting a single-run causal claim.","2026-07-28"
"BENCH-027","catalog-source-binding","As a reviewer I can prove that the runtime guidance hashed in provenance is exactly what the shipped installer generated.","Every file declared by runtime-availability.json must exist with identical bytes in the installed catalog on build and cache hit; provenance records path, SHA-256, and size.","Resolved","Commit 25a3c082; version-2 dry run ctx-ab-20260727T223036Z-1d7e43ad bound seven files and cache-hit run ctx-ab-20260727T223114Z-0be5b69e revalidated them. Adversarial packaged-copy mismatch and tampering regressions pass.","Make the benchmark hash one availability pack while the installer writes different skill bytes; catalog preparation must fail.","A cache can claim a source hash it never consumed, invalidating recommendation provenance.","Compare exact declared bytes to generated files before caching and on every cache read; version the cache schema.","Pass: the same reviewer who found the mismatch re-ran the adversarial proof and returned APPROVE with no blockers.","2026-07-28"
"BENCH-028","structured-language-retrieval","As a developer I receive language-specific runtime guidance when my host supplies language as structured context rather than repeating it in prose.","The normalized language field contributes a retrieval tag, preserves wrong-language filtering, and loads the best available language-specific skill without changing established rankings.","Resolved","Commit 48709ca8; attrs dry run ctx-ab-20260727T224932Z-326986c4 reproduced the missing skill, and ctx-ab-20260727T225220Z-f54e573f selected skill:ctx-python-testing for both identical scenarios after the fix. Local-fast then passed all 10 lanes in 44.27 seconds with 5710 unit tests passing.","Recommend for query 'fix and review frozen attrs field setter policy compatibility' with language=python, local_code_task=true, and no_api_keys=true; compare tags, results, and context_policy before and after the commit.","Ignoring structured language can hide relevant skills and make recommendation quality depend on redundant prompt wording.","Normalize the language hint once and append it to retrieval tags before ranking; retain existing availability and wrong-language filters.","Pass: independent reviewer verified graph/index paths, py canonicalization, unchanged golden rankings, and focused recommendation suites.","2026-07-28"
"BENCH-029","lifecycle-run-attestation","As a benchmark reviewer I can trust lifecycle order and know whether the committed harness and environment manifest still match at run completion.","Delivery precedes unload and session_end; the digest covers terminal events; repository HEAD/diff and the pristine in-memory environment manifest are compared at completion; any mismatch invalidates every efficiency or product claim.","Resolved","Commit 70dd4917; 94 benchmark tests and committed-head local-fast passed all 10 lanes in 47.822 seconds with 5715 unit tests passing. Corrected dry runs ctx-ab-20260727T232717Z-5c76ff5e, ctx-ab-20260727T232728Z-c94ea15c, and ctx-ab-20260727T232738Z-57479ff7 report both end attestations true and zero incidents.","Change HEAD or tracked files during a run, tamper valid environment JSON, or alter a dry-run manifest; require a run-attestation incident, nonzero exit, ineligible rows, and product_claim_eligible=false.","Startup-only cleanliness can mislabel evidence if another worker edits or commits during long trials; mutable manifests can preserve false provenance.","Retain the startup manifest in process memory, compare exact serialized bytes at completion, use precise matches-at-end wording, and fail closed for live and dry-run evidence.","Pass: independent adversarial reviewer reproduced five initial findings, verified all corrections, and returned APPROVE with no remaining findings.","2026-07-28"
"BENCH-030","deferred-activation-policy","As a host operator I can evaluate deferred MCP and agent candidates without exposing hidden evaluator data or automatically running an unsafe capability.","The pure host policy validates complete public task, activation, run, and context evidence; requires explicit false external-service, API-key, and install flags; checks provenance, permissions, and risk; selects at most one phase-appropriate capability; and otherwise denies without changing the task prompt.","Resolved","Commit d4786ccb; 116 benchmark tests passed, the adversarial reviewer returned APPROVE after closing boolean-spoofing, recursive-oracle, duplicate-ID, null-provenance, malformed-candidate, and external-install gaps, and committed-head local-fast passed all 10 lanes with 5737 tests passing in 49.163 seconds.","Pass malformed or duplicate candidates, missing or non-boolean safety flags, oracle-like nested values, null provenance, undeclared permissions, wrong-phase entity types, and valid local MCP or reviewer evidence to decide_deferred_activation.","A permissive policy could activate an external or poisoned capability, leak hidden evaluation data, or silently alter the treatment prompt.","Use strict fail-closed schemas, recursive hidden-data rejection, positive local-status allowlists, bounded permissions and risk, and immutable prompt-hash evidence.","Pass: default-deny decision semantics and adversarial boundary cases are verified; execution is tracked separately in BENCH-021.","2026-07-28"
"BENCH-031","marginal-value-abstention","As a ctx user I do not pay prompt and setup cost for guidance that only repeats the task language and adds no task-specific capability.","When a selected skill matches only the normalized scenario language and no intent-specific tag, CTX records a verified policy abstention, selects nothing, leaves the treatment prompt unchanged, and remains eligible for honest no-op comparison.","Needs Validation","Commit e8abed6f implements fail-closed language-only abstention, sealed final attestations, exact schedule enforcement, and trusted policy KPIs. Two independent reviewers approved after adversarial forged, tampered, malformed, wrong-language, constraint-only, and intent-specific cases; 135 benchmark tests passed. Dry run ctx-ab-20260728-abstention-kpi-wip confirms no selection, body fetch, prompt change, or loaded terminal state; a real current-head model turn is still required to verify the live abstention.","From a clean current committed head, run six pairs for a frozen scenario that produces an exact language-only match; require policy_abstention_verified=true, unchanged prompt hashes, zero selected and loaded context, sealed final attestations, complete exact token evidence, and no unresolved incidents.","An overbroad abstention rule could hide useful language expertise or overfit one observed benchmark; a dry run cannot prove the live model-turn evidence boundary.","Gate only an exact non-empty language-only match, fetch no body, preserve candidate and lifecycle evidence, seal results after terminal attestation, and exclude forged or untrusted rows from claims and KPI numerators.","Implementation approved with no findings; keep open only for a clean current-head live rerun and paired KPI result.","2026-07-28"
"BENCH-032","local-fast-summary-provenance","As a release reviewer I can bind every local-fast result to the exact committed head and execution window.","The summary JSON records a schema version, UTC start and finish timestamps, exact HEAD SHA, resolved merge-base SHA, profile, committed-head semantics, dirty state at selection, exact changed paths, worker count, lane timings, and return codes; every lane runs the captured HEAD.","Resolved","Commits 3f41e438 and a61237cd. Two independent reviewers approved after closing classification-before-SHA, mutable-base, ambiguous-dirty-state, and timestamp-typing findings. A clean detached-worktree run at c9e1faaba639b009c4f07e8b79e8a2dbf77e61a9 passed all 11 lanes with returncode 0 in 135.935 seconds. Ignored .gate/local-fast.json files are rolling local output and are never canonical current evidence after another run.","Run scripts/no_mistakes_run.sh fast --summary-json .gate/local-fast.json, compare head_sha to git rev-parse HEAD, verify base_sha is the merge base, require source_worktree_dirty_at_selection=false and every lane returncode=0.","Without immutable selection and execution revisions, a copied green summary or concurrent commit can misattribute lane selection and results.","Capture head and merge-base before classification, classify against those SHAs, recheck HEAD, pin every worktree to the captured revision, and serialize exact provenance.","Pass: CTO and independent reviewer approved the provenance contract; accept an individual run only when its summary binds the reviewed commit and every lane is green.","2026-07-30"
"BENCH-033","clustered-statistical-evidence","As a product owner I receive uncertainty-aware CTX benefit evidence that does not treat repeated trials as independent repositories.","Product-level verdicts collapse paired trials to scenario medians and scenarios to repository medians, require a frozen repository identity for every attempt, and use an exact repository-support test without inflating the independent sample count.","Resolved","Commit c62c81cb; 138 focused benchmark tests passed. The rolling .gate/local-fast.json binds clean head c62c81cbfcfd6a626ccbe0718f890837ee4dc5cd to 11 green lanes, 9 workers, returncode 0, and 46.587 seconds; it does not retain pytest or coverage output.","Construct three unanimous repositories, one dominant repository, missing or malformed mappings, retry-time relabeling, case-only aliases, five unanimous repositories, and a matrix with mostly verified abstentions plus one verified delivery; inspect repository_cluster_analysis, ctx_policy_outcomes, and product_benefit_verdict.","Pseudoreplication, forged repository labels, or interpreting a policy-assignment effect as universal delivered-context causality could make evidence look stronger than it is.","Require six scenarios, five canonical repositories, six trials, exact frozen URL identity on every attempt, at least one verified delivery overall, trusted delivery or verified abstention for every CTX assignment, all-repository quality and non-regression, and exact one-sided support p at most 0.05.","Pass: independent statistical and code reviewers approved the implementation; the follow-up CTO review required explicit intent-to-treat and global-delivery semantics, which are now documented without claiming per-repository delivered-context causality.","2026-07-28"
"BENCH-034","official-holdout-verification","As a benchmark reviewer I can judge baseline and CTX patches with the same hidden repository-native evaluator.","Every holdout control and scored patch runs in the pinned official SWE-bench Docker environment, proves exact FAIL_TO_PASS and PASS_TO_PASS identities, retains raw evidence, and leaves no process or container behind.","Needs Validation","HISTORICAL PLATFORM FACT: framework PR #253 merged as 0e9289e4 after all required checks passed, including the focused native Windows 3.12 gate. CURRENT POSIX CANDIDATE: the exact combined V2 suite passed 505/505 on framework code HEAD 547b80a5, and exact-head preflight passed 19/19 on 78a003ac. The isolated Python 3.12 environment and pinned Docker daemon pass dependency plus amd64/arm64 smoke checks. The official campaign remains NOT RUN: 0/10 controls, 0/30 pairs, and 0/60 arms.","From the clean merged-main worktree, authenticate the complete historical exposure inventory, run all ten controls, inspect private official reports and raw status artifacts, then execute all 60 arms and verify post-run process and container containment.","A host test path, asymmetric evaluator, unpinned image, leaked hidden artifact, or treating framework tests as production evidence could manufacture a quality result.","Use one pinned SWE-bench worker bridge for acquisition controls and both scored arms, supervise descendants and containers, authenticate every harness, runtime, dependency, image, and protocol input, and return only safe hashes and counts.","Framework integration, independent trust-boundary review, merge, and supported-host runtime conformance pass; historical Windows CI is evidence only, while the exposure inventory and official 10-control and 60-arm execution remain open.","2026-08-01"
"BENCH-035","production-graph-feature-benefit","As a product owner I can show whether the same coding agent implements real hidden features more effectively with CTX recommendations than without CTX.","A fresh frozen protocol assigns ten hidden feature tasks from ten repositories to identical baseline and CTX arms, counterbalances order, repeats pairs, keeps the evaluator secret, and reports verified quality, exact tokens, development time, errors, recommendations, and load/use/unload evidence without assuming CTX wins.","Needs Validation","CURRENT CANDIDATE: framework PR #253 merged as 0e9289e4 with all required checks green. The exact combined V2 suite passed 505/505 on framework code HEAD 547b80a5, exact-head preflight passed 19/19 on 78a003ac, and all independent review lanes returned MERGE. The official campaign remains NOT RUN: 0/10 controls, 0/30 pairs, and 0/60 arms; CTX benefit remains unproven.","For each frozen task run the same model, limits, commit, task text, evaluator, and authenticated environment in isolated baseline and CTX workspaces; execute three paired repetitions with a globally frozen 15/15 order; aggregate by repository and preserve every failure.","Small or cherry-picked tasks, different environments, leaked tests, parallel hardware contention, missing token attribution, outcome-informed reuse, or calling framework validation a product result could manufacture an apparent CTX advantage.","Use ten independent repositories, three paired repetitions per task, uncached provider tokens as the primary endpoint, development time excluding evaluator time as a secondary endpoint, quality non-inferiority, exact repository-level support, and a task-disjoint generation after any outcome-informed product fix.","Open until the complete exposure inventory, all ten controls, and 30 pairs complete with exact evidence, failures are classified, the result is independently reviewed, and the honest verdict is published. A valid not_beneficial result is final and must not be rerun or hidden.","2026-07-30"
"BENCH-036","official-v2-verdict-gates","As a benchmark reviewer I receive an official V2 verdict that cannot bypass the preregistered claim gates.","The authoritative official verdict is produced by the authenticated repository-level evaluator and requires the frozen primary token endpoint, at least nine benefiting repositories, the frozen overall token-ratio threshold, preserved quality, exact evidence, no unresolved incidents, and verified CTX delivery in all ten repositories; time-only improvement cannot pass.","Resolved","Commit 7b212141 routes official results through evaluate_repository_claim and publishes its gate results verbatim. Negative regressions reject time-only improvement and verified delivery in only nine repositories; the exact combined V2 suite passed 455/455 on HEAD 71f31993.","Construct ten repository rows with time ratio 0.80, token ratio 1.05, and verified delivery in only one repository; before the fix the generic OR-based summary could label the product beneficial, while the authenticated official evaluator now returns not_beneficial.","A favorable public verdict could contradict the frozen methodology and invalidate the confirmatory claim.","Replace the generic summary verdict for official runs with evaluate_repository_claim over authenticated repository rows, preserve every gate result in public evidence, and add negative regressions for time-only and incomplete-delivery cases.","Pass: the P0 bypass was independently reproduced, the authoritative-gate fix and regressions were reviewed, and exact-head framework integration is green; this does not claim that the official campaign has run or benefited CTX.","2026-07-30"
"BENCH-037","task-disjoint-protocol-generations","As a benchmark reviewer I can require a genuinely fresh holdout after any outcome-informed product or harness change.","Each protocol generation uses stable dataset-derived per-repository candidate ranking and selects candidate slot generation minus one; a restart either produces task-ID-disjoint choices in all ten repositories or fails closed and requires a new pinned universe.","Resolved","Commit cc1631c9 implements stable candidate partitions, candidate_slot = protocol_generation - 1, mutation rejection, and fail-closed cardinality. The methodology reviewer approved the focused 111-test evidence. The current pinned universe supports generation 1 only because 9/10 repositories have a second eligible candidate; generation 2 is therefore an intentional no-go.","Build generation 1 and generation 2 from a fixture with at least two eligible candidates per repository and compare task-ID sets, then remove one repository's second candidate or mutate the generation, seed, or slot; the first sets are disjoint and every invalid or undersupplied case is rejected.","Reusing observed tasks after a product or harness fix turns a confirmatory rerun into outcome-informed tuning and can bias the published result.","Bind generation to stable per-repository candidate slots and authenticate the complete design; after any outcome-informed fix, select a task-disjoint generation or preregister a new pinned universe when fresh cardinality is insufficient.","Pass: independent methodology review approved the task-disjoint-or-no-go design. Generation 1 remains eligible and NOT RUN; generation 2 is unavailable on the current universe by design.","2026-07-30"
"BENCH-038","authenticated-verdict-snapshots","As a benchmark reviewer I can trust that acquisition provenance and frozen verdict inputs cannot change after authentication.","The runner reconstructs the complete canonical acquisition protocol, verifies its digest and V1-derived design, retains authenticated immutable bytes for every frozen verdict input, and refuses reports or public summaries after any post-attestation drift.","Resolved","Commits fa5aa34d, 2c9b6d2b, and cd051cba authenticate canonical acquisition provenance and verdict snapshots. The independent security reviewer approved arbitrary-digest, claim-gate drift, canonical-byte, eight-input mutation, privacy, and module-invocation probes; the exact combined V2 suite passed 455/455 on HEAD 71f31993.","Supply an arbitrary acquisition digest or rehash altered claim gates, then mutate each of the eight frozen inputs after loading and request a performance report or public summary; every path is rejected before a verdict is emitted.","A forged predecessor digest or time-of-check/time-of-use mutation could produce a public verdict from unauthenticated methodology or hidden inputs.","Reconstruct and validate the canonical acquisition document, compare its canonical digest, retain authenticated bytes for all frozen artifacts, and reauthenticate snapshots at every verdict-producing boundary.","Pass: independent security review found no remaining issue in the trust-chain, post-attestation mutation, privacy, or CLI-import scope; the official campaign remains NOT RUN.","2026-07-30"
"BENCH-039","future-history-isolation","As a benchmark reviewer I can prove neither arm can inspect a future commit or gold fix from the selected repository.","Official controls and measured arms use the same authenticated offline source bundle containing the selected base commit closure, no descendant refs, no remote, and no unreachable future objects; the source-map bytes and every bundle hash are execution-frozen.","Needs Validation","Commits a3667925, e38518e7, and 65df2edb prepare, authenticate, consume, and independently revalidate exact base-closure bundles. A post-materialization future-gold bundle swap was reproduced, fixed, and independently rejected; the exact combined V2 suite passed 505/505 on code HEAD 547b80a5. Exact-head preflight then passed 19/19 on 78a003ac and PR #253 merged as 0e9289e4. HISTORICAL PLATFORM FACT: native Windows CI passed that prior candidate. No official task has been selected or run.","Create a fixture repository with a base commit followed by a future gold commit, prepare the official source artifact, clone each arm workspace, and require every ref, object, log, show, remote, and fsck probe to make the future commit inaccessible while preserving the exact base tree.","An agent can read the gold fix from local Git history, making quality, time, and token comparisons invalid even when hidden evaluator files remain private.","Retain only authenticated offline commit-closure bundles, freeze a canonical source map and bundle hashes, remove remotes after clone, and fail closed on any extra ref, unreachable object, commit, tree, or hash drift.","Implementation, adversarial tests, independent review, exact-head preflight, and merge pass on the supported contract; historical Windows CI is evidence only and authenticated live source preparation remains.","2026-08-01"
"BENCH-040","one-shot-campaign-guard","As a benchmark reviewer I can prove one frozen assignment set is measured by only one serial campaign on the host.","Official execution holds one host-wide exclusive lock for the complete campaign, independently validates frozen selection task/repository pairs against the executable scenario pack, atomically consumes their order-independent identity before measured model work, and records exact selection and execution-protocol identities as secondary fail-closed indexes, so another process, reordering, re-freeze, clone, or output path cannot overlap or replay measured assignments.","Needs Validation","The pre-launch audit first reproduced two distinct protocol digests consuming one selection on the merged framework. Independent review then reproduced reordered selection bytes producing another selection digest, a forged selection repository map producing another assignment digest for identical executable scenarios, and first-run multi-level state creation without syncing every new ancestor link. Commits 1f654635 and d4a0f8ab validate selection rules and exact scenario repositories before deriving the semantic identity, write assignment then selection then protocol indexes, fsync every verified directory/parent and each claim/containing directory even under concurrent publication, and add loader-level replay, permutation, concurrency, interruption, durability, privacy, permission, malformed-state, link-state, same-protocol, and cross-clone regressions. No official selection, protocol, assignment claim, or model outcome exists.","Claim an assignment set with selection/protocol A, then reorder or forge the selection map while retaining identical executable scenario pairs and rehash every linkage; race distinct protocols and missing hierarchy creators; inject interruption between indexes; start from a missing multi-level state base and observe every file and directory-entry fsync; every replay must fail before a model arm while one owner-only assignment record and at most one matching selection and protocol record remain.","Concurrent arms, unauthenticated semantic inputs, or byte-level reordering/re-freezing of one deterministic assignment set permits outcome-informed confirmatory reruns and invalidates timing and statistical evidence; an unsynced ancestor directory can lose every one-shot claim after a crash.","Validate fixed acquisition selection rules and exact task/repository agreement with the scenario pack, then use an owner-only persistent host lock plus an atomic semantic-assignment consumption index and exact selection/protocol secondary indexes; create and durably sync every state component and the assignment claim first, failing closed on existing, legacy, malformed, stale, aliased, mismatched, or partially written state.","Pass: five independent review cycles reproduced and closed every semantic and durability gap, ending APPROVE with no P0-P3 finding; the combined benchmark/prepare/tracker suite passed 309/309 plus Ruff, format, mypy, and diff checks; the exact committed fast gate passed all 8 lanes with 6,193 unit tests, and the activated-venv PR preflight passed all 16 steps. HISTORICAL PLATFORM FACT: native Windows CI passed before support removal. Merge to origin/main and the live one-shot campaign remain.","2026-08-01"
"BENCH-041","historical-exposure-ledger","As a benchmark reviewer I can prove every selected task is disjoint from all tasks previously shown to CTX, an LLM, or a benchmark arm.","Before ranking, selection authenticates a private canonical salted HMAC ledger, rejects every matching instance ID with explicit evidence, freezes the ledger digest into acquisition and execution protocols, and cannot proceed without the exact ledger.","Needs Validation","Commits 4b541eac, beaed567, a3667925, e38518e7, and 65df2edb bind the private HMAC ledger through selection, source preparation, controls, and freeze. Empty-ledger and overlap attacks were reproduced, fixed, and independently rejected; the exact combined V2 suite passed 505/505 and exact-head preflight passed 19/19 on 78a003ac. No V2 task was submitted to CTX or a model.","Construct a deterministic candidate set where the highest-ranked row is present in the authenticated exposure ledger; require explicit rejection and a disjoint replacement, then tamper, omit, reorder, duplicate, or swap the ledger and require fail-closed behavior.","A previously observed task allows outcome-informed tuning or memory effects and converts a confirmatory benchmark into contaminated reuse.","Build and merge a private hash-only historical ledger, authenticate exact canonical bytes, apply exclusion before all ranking/cardinality checks, and regenerate the first eligible ten-repository selection.","Implementation, independent privacy/methodology review, and exact-head preflight pass; the merged-main run must still combine all known private evidence sources and produce a fresh disjoint selection before controls.","2026-07-30"
"BENCH-042","merged-main-attestation","As a benchmark reviewer I can prove protocol creation and every measured arm use the exact clean commit published at origin/main.","The acquisition protocol freezes a credential-free origin URL and origin/main revision; protocol preparation, source preparation, startup, every arm, and final attestation require clean HEAD equal to that exact remote-tracking revision and reject drift.","Needs Validation","Commits a3667925, e38518e7, a562c443, 68d113c9, and 78a003ac authenticate origin URL, origin/main, HEAD, cleanliness, repository-state digest, and platform-native temporary-path rejection at preparation, startup, every arm, and final attestation. Exact-head preflight passed 19/19, PR #253 merged as 0e9289e4 with all required checks green, and a fresh detached worktree matched clean origin/main. Python 3.12.13 dependencies and the pinned Docker daemon passed conformance without revealing a task.","Create a clean unmerged commit, alter origin/main, change the origin URL, advance HEAD after startup, or place live scenarios under the platform temporary directory; protocol creation and each later boundary must fail before producing eligible evidence.","A clean feature branch, stale remote-tracking ref, or temporary scenario path can produce invalid or exposed evidence that appears reproducible.","Freeze sanitized remote identity and exact origin/main SHA, require HEAD equality after an authenticated fetch boundary, reject the platform temporary root, and recheck the same values before every arm and in final attestation.","Implementation, exact-head preflight, independent review, merge, and clean supported-host runtime conformance pass; historical Windows CI is evidence only and the authenticated exposure inventory remains required before selection.","2026-08-01"
"BENCH-043","preselected-descriptive-metrics","As a benchmark reviewer I receive a fixed, privacy-safe account of command, CTX funnel, lifecycle, and error evidence without changing the primary verdict after outcomes are known.","Before task selection, the runner freezes exact raw-to-aggregate definitions, retains all 60 assigned arms regardless of outcome, keeps unknown evidence missing, publishes only after the full authenticated campaign, and never feeds descriptive values into confirmatory gates.","Needs Validation","The descriptive-metrics-v1 contract and failing-first regressions cover completed shell-command calls, failures, UTF-8 output bytes, repeated commands, CTX recommendation/select/deliver/semantic-use counts, closed lifecycle events, timeouts, nonzero return codes, arm totals and medians, paired differences, four-cell binary pairs, partial-campaign withholding, duplicate rejection, sealed harness failures, primary-verdict invariance, and sentinel privacy. The official campaign remains NOT RUN: 0/10 controls, 0/30 pairs, and 0/60 arms.","Run the focused benchmark suite, mutate descriptive values while holding primary inputs fixed, inject malformed and private sentinel values, remove or duplicate an assignment, and inspect public-summary.json only after a complete 60-arm synthetic campaign.","Outcome filtering, zero-filling, calling command executions all tools, inferring semantic use from delivery, task-level publication, or post-outcome metric changes could manufacture an efficiency advantage.","Commit the versioned descriptive schema before selection; validate complete traces and closed lifecycles; report known and missing denominators; use fixed whole-arm and whole-campaign aggregates; enforce an allowlist-only public boundary; leave primary thresholds and verdict code unchanged.","Implementation tests pass locally; independent code and methodology review plus required CI remain before any task selection or production arm.","2026-08-01"