feat: update README and runbook for SFT and GRPO training enhancements

Revised the README.md and runbook to reflect changes in the SFT and GRPO training processes, including default input/output paths and parameters. Updated example commands for data preparation and training scripts to incorporate new defaults for merging datasets. Added details on the new `e1_m1_h1_examples.jsonl` file and clarified the training steps for improved user guidance. Introduced new JSONL files for verification results and summaries to enhance tracking of model performance.

README.md

@@ -71,9 +71,11 @@ Replay generated examples against the current environment before training:
 
 ```powershell
 uv run python scripts/verify_examples.py --input data/raw/e1_m1_examples.jsonl --clean data/processed/e1_m1_clean.jsonl --rejected data/processed/e1_m1_rejected.jsonl
-uv run python scripts/prepare_sft_data.py --input data/processed/e1_m1_clean.jsonl --output data/sft/e1_m1_examples.jsonl
+uv run python scripts/prepare_sft_data.py
 ```
 
+`prepare_sft_data.py` defaults to merging `data/processed/e1_m1_clean.jsonl` and `data/processed/h1_seed_clean.jsonl` into `data/sft/e1_m1_h1_examples.jsonl`. To build E/M only, pass e.g. `--input data/processed/e1_m1_clean.jsonl --output data/sft/e1_m1_examples.jsonl`.
+
 Generate a small H1 seed set if needed:
 
 ```powershell
@@ -87,10 +89,12 @@ Training scripts use **Unsloth** and **Hugging Face TRL** and are intended for a
 
 ```bash
 pip install -e ".[training]"
-python training/train_sft.py --model Qwen/Qwen2.5-7B-Instruct --data data/sft/e1_m1_examples.jsonl --output outputs/sft_adapter --max-steps 30
-python training/train_grpo.py --model Qwen/Qwen2.5-7B-Instruct --adapter outputs/sft_adapter --examples data/processed/e1_m1_clean.jsonl --output outputs/grpo_adapter --max-steps 30
+python training/train_sft.py --model Qwen/Qwen2.5-7B-Instruct --data data/sft/e1_m1_h1_examples.jsonl --output outputs/sft_adapter --max-steps 30
+python training/train_grpo.py --model Qwen/Qwen2.5-7B-Instruct --adapter outputs/sft_adapter --output outputs/grpo_adapter --max-steps 30
 ```
 
+`train_grpo.py` defaults to `--max-steps 150`, E/M/H task mix, and comma-separated `--examples` including H1 when those files exist. For a short smoke test, add e.g. `--max-steps 30`.
+
 For a judge-rerunnable notebook, use [`notebooks/corp_env_trl_unsloth_training.ipynb`](notebooks/corp_env_trl_unsloth_training.ipynb).
 
 See [`docs/lightning_hf_runbook.md`](docs/lightning_hf_runbook.md) for the short-session H100 + Hugging Face workflow.

data/raw/h1_seed.jsonl

@@ -1,8 +1,24 @@
 {"example_id": "seed-h1_acquisition_defence-000", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-001", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-002", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-003", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-004", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-005", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-006", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
-{"example_id": "seed-h1_acquisition_defence-007", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-001", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Map product/IP defensibility and engineering dependency risk in the target."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Model valuation bands against board-approved cash and runway headroom."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Quantify people flight risk and retention levers for critical engineering staff."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.\", \"deadline\": \"Cap negotiations at 50 days: balance hr timelines with cash runway and burn.\", \"retention_plan\": \"Retention bonuses plus explicit reporting lines for VPs in core product.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-002", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Report whether technical differentiation justifies a premium vs the 2.3x bid."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Clarify how many months of runway remain if we stretch price or drag timing."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Clarify CHRO's view on 90-day people risk and hiring-market pressure."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Signal 3.25x as the opening position with board ratified walk-down rights.\", \"deadline\": \"Close or exit talks within 40 days; 7 month runway does not allow drift.\", \"retention_plan\": \"Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-003", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Synthesize the CTO view on product moat and technical integration costs."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Synthesize the CFO view on the ceiling the board can defend with current cash."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Synthesize the CHRO view on retention, titles, and cultural integration risk."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.\", \"deadline\": \"45-day triage: align cfo cash tests with cto and chro risk windows.\", \"retention_plan\": \"Milestone-based retention: pay at signing, 90d, and close for key dev talent.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-004", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: outline minimum acceptable tech valuation given integration complexity."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: outline a pricing corridor consistent with 7 month runway and board guardrails."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: outline retention and communication needs if diligence slips past the HR window."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-005", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: stress-test whether acquirer can credibly match our roadmap without the team."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: frame downside if we overpay vs observed comps and our burn profile."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: flag key engineers whose departure would void strategic upside from the deal."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.\", \"deadline\": \"Cap negotiations at 50 days: balance hr timelines with cash runway and burn.\", \"retention_plan\": \"Retention bonuses plus explicit reporting lines for VPs in core product.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-006", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Signal 3.25x as the opening position with board ratified walk-down rights.\", \"deadline\": \"Close or exit talks within 40 days; 7 month runway does not allow drift.\", \"retention_plan\": \"Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-007", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Map product/IP defensibility and engineering dependency risk in the target."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Model valuation bands against board-approved cash and runway headroom."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Quantify people flight risk and retention levers for critical engineering staff."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.\", \"deadline\": \"45-day triage: align cfo cash tests with cto and chro risk windows.\", \"retention_plan\": \"Milestone-based retention: pay at signing, 90d, and close for key dev talent.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-008", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Report whether technical differentiation justifies a premium vs the 2.3x bid."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Clarify how many months of runway remain if we stretch price or drag timing."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Clarify CHRO's view on 90-day people risk and hiring-market pressure."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-009", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Synthesize the CTO view on product moat and technical integration costs."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Synthesize the CFO view on the ceiling the board can defend with current cash."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Synthesize the CHRO view on retention, titles, and cultural integration risk."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.\", \"deadline\": \"Cap negotiations at 50 days: balance hr timelines with cash runway and burn.\", \"retention_plan\": \"Retention bonuses plus explicit reporting lines for VPs in core product.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-010", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: outline minimum acceptable tech valuation given integration complexity."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: outline a pricing corridor consistent with 7 month runway and board guardrails."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: outline retention and communication needs if diligence slips past the HR window."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Signal 3.25x as the opening position with board ratified walk-down rights.\", \"deadline\": \"Close or exit talks within 40 days; 7 month runway does not allow drift.\", \"retention_plan\": \"Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-011", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: stress-test whether acquirer can credibly match our roadmap without the team."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: frame downside if we overpay vs observed comps and our burn profile."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: flag key engineers whose departure would void strategic upside from the deal."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.\", \"deadline\": \"45-day triage: align cfo cash tests with cto and chro risk windows.\", \"retention_plan\": \"Milestone-based retention: pay at signing, 90d, and close for key dev talent.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-012", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-013", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Map product/IP defensibility and engineering dependency risk in the target."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Model valuation bands against board-approved cash and runway headroom."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Quantify people flight risk and retention levers for critical engineering staff."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.\", \"deadline\": \"Cap negotiations at 50 days: balance hr timelines with cash runway and burn.\", \"retention_plan\": \"Retention bonuses plus explicit reporting lines for VPs in core product.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-014", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Report whether technical differentiation justifies a premium vs the 2.3x bid."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Clarify how many months of runway remain if we stretch price or drag timing."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Clarify CHRO's view on 90-day people risk and hiring-market pressure."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Signal 3.25x as the opening position with board ratified walk-down rights.\", \"deadline\": \"Close or exit talks within 40 days; 7 month runway does not allow drift.\", \"retention_plan\": \"Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-015", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Synthesize the CTO view on product moat and technical integration costs."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Synthesize the CFO view on the ceiling the board can defend with current cash."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Synthesize the CHRO view on retention, titles, and cultural integration risk."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.\", \"deadline\": \"45-day triage: align cfo cash tests with cto and chro risk windows.\", \"retention_plan\": \"Milestone-based retention: pay at signing, 90d, and close for key dev talent.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-016", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: outline minimum acceptable tech valuation given integration complexity."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: outline a pricing corridor consistent with 7 month runway and board guardrails."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: outline retention and communication needs if diligence slips past the HR window."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-017", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: stress-test whether acquirer can credibly match our roadmap without the team."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: frame downside if we overpay vs observed comps and our burn profile."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: flag key engineers whose departure would void strategic upside from the deal."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.\", \"deadline\": \"Cap negotiations at 50 days: balance hr timelines with cash runway and burn.\", \"retention_plan\": \"Retention bonuses plus explicit reporting lines for VPs in core product.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-018", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Signal 3.25x as the opening position with board ratified walk-down rights.\", \"deadline\": \"Close or exit talks within 40 days; 7 month runway does not allow drift.\", \"retention_plan\": \"Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-019", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Map product/IP defensibility and engineering dependency risk in the target."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Model valuation bands against board-approved cash and runway headroom."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Quantify people flight risk and retention levers for critical engineering staff."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.\", \"deadline\": \"45-day triage: align cfo cash tests with cto and chro risk windows.\", \"retention_plan\": \"Milestone-based retention: pay at signing, 90d, and close for key dev talent.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-020", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Report whether technical differentiation justifies a premium vs the 2.3x bid."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Clarify how many months of runway remain if we stretch price or drag timing."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Clarify CHRO's view on 90-day people risk and hiring-market pressure."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Open at 3.2x with a board-approved walk-down floor near 2.6x.\", \"deadline\": \"Force a decision inside 45 days to preserve cash runway and reduce retention risk.\", \"retention_plan\": \"Offer retention grants and role clarity for critical engineering leaders.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-021", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "Synthesize the CTO view on product moat and technical integration costs."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "Synthesize the CFO view on the ceiling the board can defend with current cash."}, {"action_type": "delegate", "agent_id": "chro", "payload": "Synthesize the CHRO view on retention, titles, and cultural integration risk."}, {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO valuation ambition conflicts with CFO runway and cash constraints.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A slow process increases CHRO retention risk for key engineering talent.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Ask above the CFO ceiling but set a fast deadline and walk-down logic.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."}, {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.\", \"deadline\": \"Cap negotiations at 50 days: balance hr timelines with cash runway and burn.\", \"retention_plan\": \"Retention bonuses plus explicit reporting lines for VPs in core product.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-022", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: outline minimum acceptable tech valuation given integration complexity."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: outline a pricing corridor consistent with 7 month runway and board guardrails."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: outline retention and communication needs if diligence slips past the HR window."}, {"action_type": "log_reasoning", "payload": "The cto case implies we should not fold to 2.3x without a premium for IP."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"CTO's strategic premium clashes with cfo and finance on cash runway and board limits.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "Finance and cash runway cap how long we can posture before credibility breaks."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"A drawn-out haggle raises chro and hr risk as competing offers come due.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "A 7 month cash runway means every extra week in diligence burns optionality and cash."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Choose a path that fits both the 7 month cash runway and board finance constraints."}, {"action_type": "log_reasoning", "payload": "chro: retention grants must land before the team interprets delay as a leadership stall."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Signal 3.25x as the opening position with board ratified walk-down rights.\", \"deadline\": \"Close or exit talks within 40 days; 7 month runway does not allow drift.\", \"retention_plan\": \"Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.\"}"}]}
+{"example_id": "seed-h1_acquisition_defence-023", "task_id": "h1_acquisition_defence", "source": "scripted_seed", "actions": [{"action_type": "delegate", "agent_id": "cto", "payload": "CTO: stress-test whether acquirer can credibly match our roadmap without the team."}, {"action_type": "delegate", "agent_id": "cfo", "payload": "CFO: frame downside if we overpay vs observed comps and our burn profile."}, {"action_type": "delegate", "agent_id": "chro", "payload": "CHRO: flag key engineers whose departure would void strategic upside from the deal."}, {"action_type": "log_reasoning", "payload": "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c1\", \"summary\": \"Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.\", \"source_agents\": [\"cto\", \"cfo\"]}"}, {"action_type": "log_reasoning", "payload": "CFO: board cannot fund an endless auction with only about 7 month runway in reserve."}, {"action_type": "log_conflict", "payload": "{\"id\": \"c2\", \"summary\": \"chro and cto both warn: delay past the HR window erodes the engineering team.\", \"source_agents\": [\"chro\", \"cto\"]}"}, {"action_type": "log_resolution", "payload": "{\"conflict_id\": \"c1\", \"resolution_type\": \"bounded_counter\", \"text\": \"Use a time-boxed process: name a number above cfo\\u2019s ceiling, then step down on a defined clock.\"}"}, {"action_type": "advance_phase", "payload": "analysis"}, {"action_type": "log_reasoning", "payload": "Runway and cash pressure force a firm deadline, not an open-ended beauty contest."}, {"action_type": "advance_phase", "payload": "decision"}, {"action_type": "log_decision", "payload": "Commit: respect runway and cash reality while still pressing for a fair tech premium."}, {"action_type": "log_reasoning", "payload": "hr signals that clarity on roles and comp must precede a public counter narrative."}, {"action_type": "advance_phase", "payload": "execution"}, {"action_type": "log_reasoning", "payload": "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package."}, {"action_type": "finalize", "payload": "{\"counter_offer\": \"Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.\", \"deadline\": \"45-day triage: align cfo cash tests with cto and chro risk windows.\", \"retention_plan\": \"Milestone-based retention: pay at signing, 90d, and close for key dev talent.\"}"}]}

docs/lightning_hf_runbook.md

@@ -1,15 +1,15 @@
 # Lightning AI + Hugging Face Runbook
 
-This runbook is optimized for short 3-4 hour H100 windows and Hugging Face credits. The judge-rerunnable notebook version is [`notebooks/corp_env_trl_unsloth_training.ipynb`](../notebooks/corp_env_trl_unsloth_training.ipynb).
+This runbook is optimized for short 3-4 hour H100 windows and Hugging Face credits. The judge-rerunnable notebook version is [`notebooks/corp_env_trl_unsloth_training.ipynb`](../notebooks/corp_env_trl_unsloth_training.ipynb). A minified (single-line) Colab export in the repo can be searched for `train_grpo` or `MAX_STEPS` to recover pins or a prior command line, then align with the defaults in `training/train_grpo.py` and `training/train_sft.py`.
 
 ## 1. Prepare Artifacts Locally
 
 ```powershell
 uv sync --extra dev --extra plots
-uv run python scripts/generate_sft_data.py --tasks h1_acquisition_defence --per-task 8 --output data/raw/h1_seed.jsonl
+uv run python scripts/generate_sft_data.py --tasks h1_acquisition_defence --per-task 24 --output data/raw/h1_seed.jsonl
 uv run python scripts/verify_examples.py --input data/raw/e1_m1_examples.jsonl --clean data/processed/e1_m1_clean.jsonl --rejected data/processed/e1_m1_rejected.jsonl
 uv run python scripts/verify_examples.py --input data/raw/h1_seed.jsonl --clean data/processed/h1_seed_clean.jsonl --rejected data/processed/h1_seed_rejected.jsonl
-uv run python scripts/prepare_sft_data.py --input data/processed/e1_m1_clean.jsonl --output data/sft/e1_m1_examples.jsonl
+uv run python scripts/prepare_sft_data.py
 ```
 
 Keep raw examples untouched. Train only from verified `data/processed/*_clean.jsonl` and `data/sft/*.jsonl`.
@@ -19,11 +19,12 @@ Keep raw examples untouched. Train only from verified `data/processed/*_clean.js
 ```powershell
 huggingface-cli login
 huggingface-cli repo create corp-env-data --type dataset
-huggingface-cli upload <your-user-or-org>/corp-env-data data/sft/e1_m1_examples.jsonl data/sft/e1_m1_examples.jsonl --repo-type dataset
+huggingface-cli upload <your-user-or-org>/corp-env-data data/sft/e1_m1_h1_examples.jsonl data/sft/e1_m1_h1_examples.jsonl --repo-type dataset
 huggingface-cli upload <your-user-or-org>/corp-env-data data/processed/e1_m1_clean.jsonl data/processed/e1_m1_clean.jsonl --repo-type dataset
+huggingface-cli upload <your-user-or-org>/corp-env-data data/processed/h1_seed_clean.jsonl data/processed/h1_seed_clean.jsonl --repo-type dataset
 ```
 
-Also upload `data/processed/h1_seed_clean.jsonl` if the H1 seed passes verification.
+Retire or mirror `e1_m1_examples.jsonl` in the same dataset if you need backward-compatible paths.
 
 ## 3. Lightning H100 Session 1: SFT
 
@@ -37,14 +38,14 @@ pip install -e ".[training]"
 huggingface-cli login
 python training/train_sft.py \
   --model Qwen/Qwen2.5-7B-Instruct \
-  --data data/sft/e1_m1_examples.jsonl \
+  --data data/sft/e1_m1_h1_examples.jsonl \
   --output outputs/sft_adapter \
   --epochs 2 \
   --max-steps 30 \
   --push-to-hub <your-user-or-org>/corp-env-sft-adapter
 ```
 
-This uses Unsloth + TRL `SFTTrainer`. If setup time is short, use a 7B model. If the session is stable and examples are clean, try a 14B model for the SFT demo. Remove `--max-steps 30` or raise it for a real run.
+This uses Unsloth + TRL `SFTTrainer` with conversational `messages` JSONL (and optional `--packing` / `--dataloader-num-workers` on a strong box). If setup time is short, use a 7B model. If the session is stable and examples are clean, try a 14B model for the SFT demo. Remove `--max-steps 30` or raise it for a real run. Install a `flash-attn` wheel that matches the session’s `torch`+CUDA for best GRPO and long-context SFT step time.
 
 ## 4. Lightning H100 Session 2: Eval SFT
 
@@ -66,14 +67,14 @@ Push `results/*.jsonl` to Hugging Face or copy them back before the Lightning ac
 python training/train_grpo.py \
   --model Qwen/Qwen2.5-7B-Instruct \
   --adapter outputs/sft_adapter \
-  --examples data/processed/e1_m1_clean.jsonl \
+  --examples data/processed/e1_m1_clean.jsonl,data/processed/h1_seed_clean.jsonl \
   --output outputs/grpo_adapter \
-  --tasks e1_launch_readiness,m1_budget_reallocation \
-  --max-steps 30 \
+  --max-steps 150 \
+  --dataloader-num-workers 2 \
   --push-to-hub <your-user-or-org>/corp-env-grpo-adapter
 ```
 
-This uses Unsloth + TRL `GRPOTrainer` with the real CORP-ENV reward path. Start with E1/M1. Add H1 only after E1/M1 rewards are non-zero and invalid action rate is low. For a real run, increase `--max-steps` to `150` or more.
+This uses Unsloth + TRL `GRPOTrainer` with the real CORP-ENV reward path. Defaults now include H1 in `--tasks` and a 150-step run; pass `--max-steps 30` for smoke only. For throughput on H100, match FlashAttention to your `torch` build, then try larger `--batch-size` or more `--dataloader-num-workers` if the CPU and dataset I/O allow.
 
 ## 6. Final Eval And Plots
 

notebooks/corp_env_trl_unsloth_training.ipynb

@@ -1,252 +1,253 @@
 {
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# CORP-ENV TRL + Unsloth Training Notebook\n",
-    "\n",
-    "This notebook is the judge-rerunnable training path for CORP-ENV. It verifies examples, prepares SFT data, runs Unsloth + TRL SFT, optionally runs Unsloth + TRL GRPO, evaluates model stages through the OpenEnv environment, and generates result plots.\n",
-    "\n",
-    "Recommended runtime: Colab GPU, Lightning AI H100, or another CUDA GPU machine. For a quick judge smoke run, keep `MAX_STEPS` small."
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 1. Setup\n",
-    "\n",
-    "In Colab, set `REPO_URL` to your GitHub or Hugging Face repo URL. If the repo is already mounted/cloned, skip the clone cell and `cd` into the repo."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "REPO_URL = \"https://github.com/<your-user>/<your-repo>.git\"  # TODO: replace\n",
-    "BASE_MODEL = \"Qwen/Qwen2.5-7B-Instruct\"\n",
-    "HF_ORG_OR_USER = \"<your-user-or-org>\"  # TODO: replace\n",
-    "MAX_STEPS = 30  # quick judge smoke; increase for real training\n",
-    "RUN_GRPO = False  # set True after SFT smoke passes"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!git clone {REPO_URL} corp_gym || true\n",
-    "%cd corp_gym\n",
-    "!pip install -U pip\n",
-    "!pip install -e \".[training,plots]\""
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 2. Optional Hugging Face Login\n",
-    "\n",
-    "Run this if you want to download private models/datasets or push adapters."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from huggingface_hub import notebook_login\n",
-    "notebook_login()"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 3. Local OpenEnv Checks\n",
-    "\n",
-    "These are lightweight and should pass before training."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!python -m unittest discover -s tests\n",
-    "!openenv validate"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 4. Verify Examples And Build SFT Data\n",
-    "\n",
-    "If `data/raw/e1_m1_examples.jsonl` already exists, the import command is harmless to skip. The verification step replays examples through `CorpEnvironment` and filters bad trajectories."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!python scripts/import_generated_examples.py \\\n",
-    "  --inputs data/raw/e1_to_e100_tasks.py data/raw/m1_to_m100_tasks.py \\\n",
-    "  --output data/raw/e1_m1_examples.jsonl\n",
-    "\n",
-    "!python scripts/verify_examples.py \\\n",
-    "  --input data/raw/e1_m1_examples.jsonl \\\n",
-    "  --clean data/processed/e1_m1_clean.jsonl \\\n",
-    "  --rejected data/processed/e1_m1_rejected.jsonl \\\n",
-    "  --summary results/e1_m1_verification_summary.json \\\n",
-    "  --all-records results/e1_m1_verification_all.jsonl\n",
-    "\n",
-    "!python scripts/prepare_sft_data.py \\\n",
-    "  --input data/processed/e1_m1_clean.jsonl \\\n",
-    "  --output data/sft/e1_m1_examples.jsonl"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 5. Baseline Eval\n",
-    "\n",
-    "This gives a pre-training comparison point."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!python eval.py --policy scripted_weak --label baseline --output results/baseline_eval.jsonl\n",
-    "!python eval.py --policy oracle --label oracle --output results/oracle_eval.jsonl"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 6. SFT With Unsloth + TRL\n",
-    "\n",
-    "This runs `training/train_sft.py`, which uses `unsloth.FastLanguageModel` and `trl.SFTTrainer`. For a real run, increase `MAX_STEPS` or remove the flag and use epochs."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!python training/train_sft.py \\\n",
-    "  --model {BASE_MODEL} \\\n",
-    "  --data data/sft/e1_m1_examples.jsonl \\\n",
-    "  --output outputs/sft_adapter \\\n",
-    "  --max-steps {MAX_STEPS} \\\n",
-    "  --push-to-hub {HF_ORG_OR_USER}/corp-env-sft-adapter"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 7. Evaluate SFT Adapter"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!python eval.py \\\n",
-    "  --policy hf \\\n",
-    "  --label sft \\\n",
-    "  --model {BASE_MODEL} \\\n",
-    "  --adapter outputs/sft_adapter \\\n",
-    "  --output results/sft_eval.jsonl"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 8. Optional GRPO With Unsloth + TRL\n",
-    "\n",
-    "This runs `training/train_grpo.py`, which uses `trl.GRPOTrainer` and calls the real CORP-ENV environment reward. Keep this off for the first smoke run."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "if RUN_GRPO:\n",
-    "    !python training/train_grpo.py \\\n",
-    "      --model {BASE_MODEL} \\\n",
-    "      --adapter outputs/sft_adapter \\\n",
-    "      --examples data/processed/e1_m1_clean.jsonl \\\n",
-    "      --output outputs/grpo_adapter \\\n",
-    "      --tasks e1_launch_readiness,m1_budget_reallocation \\\n",
-    "      --max-steps {MAX_STEPS} \\\n",
-    "      --push-to-hub {HF_ORG_OR_USER}/corp-env-grpo-adapter\n",
-    "\n",
-    "    !python eval.py \\\n",
-    "      --policy hf \\\n",
-    "      --label grpo \\\n",
-    "      --model {BASE_MODEL} \\\n",
-    "      --adapter outputs/grpo_adapter \\\n",
-    "      --output results/grpo_eval.jsonl"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## 9. Generate Final Plots"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "inputs = \"results/baseline_eval.jsonl results/sft_eval.jsonl\"\n",
-    "if RUN_GRPO:\n",
-    "    inputs += \" results/grpo_eval.jsonl\"\n",
-    "!python plot_results.py --inputs {inputs} --output-dir results\n",
-    "!ls -lh results"
-   ]
-  }
- ],
- "metadata": {
-  "accelerator": "GPU",
-  "colab": {
-   "gpuType": "T4",
-   "provenance": []
-  },
-  "kernelspec": {
-   "display_name": "Python 3",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "name": "python",
-   "pygments_lexer": "ipython3"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 5
-}
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "# CORP-ENV TRL + Unsloth Training Notebook\n",
+        "\n",
+        "This notebook is the judge-rerunnable training path for CORP-ENV. It verifies examples, prepares SFT data, runs Unsloth + TRL SFT, optionally runs Unsloth + TRL GRPO, evaluates model stages through the OpenEnv environment, and generates result plots.\n",
+        "\n",
+        "Recommended runtime: Colab GPU, Lightning AI H100, or another CUDA GPU machine. For a quick judge smoke run, keep `MAX_STEPS` small."
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 1. Setup\n",
+        "\n",
+        "In Colab, set `REPO_URL` to your GitHub or Hugging Face repo URL. If the repo is already mounted/cloned, skip the clone cell and `cd` into the repo."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "REPO_URL = \"https://github.com/<your-user>/<your-repo>.git\"  # TODO: replace\n",
+        "BASE_MODEL = \"Qwen/Qwen2.5-7B-Instruct\"\n",
+        "HF_ORG_OR_USER = \"<your-user-or-org>\"  # TODO: replace\n",
+        "MAX_STEPS = 30  # SFT quick judge smoke; remove or lower for a full-epoch SFT run\n",
+        "GRPO_MAX_STEPS = 150  # used only when RUN_GRPO is True\n",
+        "RUN_GRPO = False  # set True after SFT smoke passes"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "!git clone {REPO_URL} corp_gym || true\n",
+        "%cd corp_gym\n",
+        "!pip install -U pip\n",
+        "!pip install -e \".[training,plots]\""
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 2. Optional Hugging Face Login\n",
+        "\n",
+        "Run this if you want to download private models/datasets or push adapters."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "from huggingface_hub import notebook_login\n",
+        "notebook_login()"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 3. Local OpenEnv Checks\n",
+        "\n",
+        "These are lightweight and should pass before training."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "!python -m unittest discover -s tests\n",
+        "!openenv validate"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 4. Verify Examples And Build SFT Data\n",
+        "\n",
+        "If `data/raw/e1_m1_examples.jsonl` already exists, the import command is harmless to skip. The verification step replays examples through `CorpEnvironment` and filters bad trajectories."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "!python scripts/import_generated_examples.py \\\n",
+        "  --inputs data/raw/e1_to_e100_tasks.py data/raw/m1_to_m100_tasks.py \\\n",
+        "  --output data/raw/e1_m1_examples.jsonl\n",
+        "\n",
+        "!python scripts/verify_examples.py \\\n",
+        "  --input data/raw/e1_m1_examples.jsonl \\\n",
+        "  --clean data/processed/e1_m1_clean.jsonl \\\n",
+        "  --rejected data/processed/e1_m1_rejected.jsonl \\\n",
+        "  --summary results/e1_m1_verification_summary.json \\\n",
+        "  --all-records results/e1_m1_verification_all.jsonl\n",
+        "\n",
+        "!python scripts/generate_sft_data.py --tasks h1_acquisition_defence --per-task 24 --output data/raw/h1_seed.jsonl\n",
+        "!python scripts/verify_examples.py --input data/raw/h1_seed.jsonl --clean data/processed/h1_seed_clean.jsonl --rejected data/processed/h1_seed_rejected.jsonl\n",
+        "!python scripts/prepare_sft_data.py"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 5. Baseline Eval\n",
+        "\n",
+        "This gives a pre-training comparison point."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "!python eval.py --policy scripted_weak --label baseline --output results/baseline_eval.jsonl\n",
+        "!python eval.py --policy oracle --label oracle --output results/oracle_eval.jsonl"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 6. SFT With Unsloth + TRL\n",
+        "\n",
+        "This runs `training/train_sft.py`, which uses `unsloth.FastLanguageModel` and `trl.SFTTrainer`. For a real run, increase `MAX_STEPS` or remove the flag and use epochs."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "!python training/train_sft.py \\\n",
+        "  --model {BASE_MODEL} \\\n",
+        "  --data data/sft/e1_m1_h1_examples.jsonl \\\n",
+        "  --output outputs/sft_adapter \\\n",
+        "  --max-steps {MAX_STEPS} \\\n",
+        "  --push-to-hub {HF_ORG_OR_USER}/corp-env-sft-adapter"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 7. Evaluate SFT Adapter"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "!python eval.py \\\n",
+        "  --policy hf \\\n",
+        "  --label sft \\\n",
+        "  --model {BASE_MODEL} \\\n",
+        "  --adapter outputs/sft_adapter \\\n",
+        "  --output results/sft_eval.jsonl"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 8. Optional GRPO With Unsloth + TRL\n",
+        "\n",
+        "This runs `training/train_grpo.py`, which uses `trl.GRPOTrainer` and calls the real CORP-ENV environment reward. Keep this off for the first smoke run."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "if RUN_GRPO:\n",
+        "    !python training/train_grpo.py \\\n",
+        "      --model {BASE_MODEL} \\\n",
+        "      --adapter outputs/sft_adapter \\\n",
+        "      --examples data/processed/e1_m1_clean.jsonl,data/processed/h1_seed_clean.jsonl \\\n",
+        "      --output outputs/grpo_adapter \\\n",
+        "      --max-steps {GRPO_MAX_STEPS} \\\n",
+        "      --dataloader-num-workers 2 \\\n",
+        "      --push-to-hub {HF_ORG_OR_USER}/corp-env-grpo-adapter\n",
+        "\n",
+        "    !python eval.py \\\n",
+        "      --policy hf \\\n",
+        "      --label grpo \\\n",
+        "      --model {BASE_MODEL} \\\n",
+        "      --adapter outputs/grpo_adapter \\\n",
+        "      --output results/grpo_eval.jsonl"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## 9. Generate Final Plots"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "inputs = \"results/baseline_eval.jsonl results/sft_eval.jsonl\"\n",
+        "if RUN_GRPO:\n",
+        "    inputs += \" results/grpo_eval.jsonl\"\n",
+        "!python plot_results.py --inputs {inputs} --output-dir results\n",
+        "!ls -lh results"
+      ],
+      "execution_count": null,
+      "outputs": []
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "T4",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python",
+      "pygments_lexer": "ipython3"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 5
+}

results/example_verification_summary.json

@@ -0,0 +1,16 @@
+{
+  "total": 24,
+  "by_status": {
+    "clean": 24
+  },
+  "by_reject_reason": {
+    "clean": 24
+  },
+  "by_task": {
+    "h1_acquisition_defence": {
+      "clean": 24
+    }
+  },
+  "clean_avg_terminal_reward": 1.05,
+  "clean_avg_verifier_pass_rate": 1.0
+}

scripts/_trajectory_utils.py

@@ -383,33 +383,134 @@ def oracle_actions(task_id: str, variant: int = 0) -> List[Dict[str, Any]]:
             {"action_type": "finalize", "payload": json.dumps(phase)},
         ]
     if task_id == "h1_acquisition_defence":
-        final = {
-            "counter_offer": "Open at 3.2x with a board-approved walk-down floor near 2.6x.",
-            "deadline": "Force a decision inside 45 days to preserve cash runway and reduce retention risk.",
-            "retention_plan": "Offer retention grants and role clarity for critical engineering leaders.",
-        }
+        # Rotating templates: paraphrase payloads while preserving verifier invariants
+        # (3 delegates, 2 conflicts, resolution_type, decisions mentioning runway/cash/7 month,
+        # final dict keys, phase progression, dense reasoning).
+        n = 12
+        v = variant % n
+        delegates = [
+            (
+                "Assess IP leverage and technical moat.",
+                "Assess cash runway and valuation ceiling.",
+                "Assess retention risk and people constraints.",
+            ),
+            (
+                "Map product/IP defensibility and engineering dependency risk in the target.",
+                "Model valuation bands against board-approved cash and runway headroom.",
+                "Quantify people flight risk and retention levers for critical engineering staff.",
+            ),
+            (
+                "Report whether technical differentiation justifies a premium vs the 2.3x bid.",
+                "Clarify how many months of runway remain if we stretch price or drag timing.",
+                "Clarify CHRO's view on 90-day people risk and hiring-market pressure.",
+            ),
+            (
+                "Synthesize the CTO view on product moat and technical integration costs.",
+                "Synthesize the CFO view on the ceiling the board can defend with current cash.",
+                "Synthesize the CHRO view on retention, titles, and cultural integration risk.",
+            ),
+            (
+                "CTO: outline minimum acceptable tech valuation given integration complexity.",
+                "CFO: outline a pricing corridor consistent with 7 month runway and board guardrails.",
+                "CHRO: outline retention and communication needs if diligence slips past the HR window.",
+            ),
+            (
+                "CTO: stress-test whether acquirer can credibly match our roadmap without the team.",
+                "CFO: frame downside if we overpay vs observed comps and our burn profile.",
+                "CHRO: flag key engineers whose departure would void strategic upside from the deal.",
+            ),
+        ]
+        d0, d1, d2 = delegates[v % len(delegates)]
+        c1_sums = [
+            "CTO valuation ambition conflicts with CFO runway and cash constraints.",
+            "CTO's strategic premium clashes with cfo and finance on cash runway and board limits.",
+            "Engineering (cto) wants a high counter; finance (cfo) caps what cash runway allows.",
+        ]
+        c2_sums = [
+            "A slow process increases CHRO retention risk for key engineering talent.",
+            "A drawn-out haggle raises chro and hr risk as competing offers come due.",
+            "chro and cto both warn: delay past the HR window erodes the engineering team.",
+        ]
+        res_texts = [
+            "Ask above the CFO ceiling but set a fast deadline and walk-down logic.",
+            "Pursue a structured counter: higher opening ask with a board-mandated walk-down to 2.6x floor.",
+            "Use a time-boxed process: name a number above cfo’s ceiling, then step down on a defined clock.",
+        ]
+        r_before_a = [
+            "CTO input supports a higher counter because the IP moat is meaningful.",
+            "The cto case implies we should not fold to 2.3x without a premium for IP.",
+            "Dev leadership sees differentiated tech, so a weak counter cedes leverage too early.",
+        ]
+        r_after_c1 = [
+            "CFO input limits how long the negotiation can remain open.",
+            "Finance and cash runway cap how long we can posture before credibility breaks.",
+            "CFO: board cannot fund an endless auction with only about 7 month runway in reserve.",
+        ]
+        r_runway = [
+            "The 7 month runway makes delay costly and should shape the deadline.",
+            "A 7 month cash runway means every extra week in diligence burns optionality and cash.",
+            "Runway and cash pressure force a firm deadline, not an open-ended beauty contest.",
+        ]
+        r_retention = [
+            "Retention incentives are required before the market reads uncertainty.",
+            "chro: retention grants must land before the team interprets delay as a leadership stall.",
+            "hr signals that clarity on roles and comp must precede a public counter narrative.",
+        ]
+        r_final = [
+            "Final recommendation integrates valuation, deadline, and retention plan.",
+            "Synthesize cto, cfo, and chro inputs into one executable counter, timeline, and people plan.",
+            "Close the loop: a bounded counter, a calendar-driven deadline, and a retention package.",
+        ]
+        decision_line = [
+            "Proceed with a counter that acknowledges 7 month runway and cash limits.",
+            "Choose a path that fits both the 7 month cash runway and board finance constraints.",
+            "Commit: respect runway and cash reality while still pressing for a fair tech premium.",
+        ]
+        finals = [
+            {
+                "counter_offer": "Open at 3.2x with a board-approved walk-down floor near 2.6x.",
+                "deadline": "Force a decision inside 45 days to preserve cash runway and reduce retention risk.",
+                "retention_plan": "Offer retention grants and role clarity for critical engineering leaders.",
+            },
+            {
+                "counter_offer": "Start at 3.15x with a staged retreat toward 2.65x if diligence stays clean.",
+                "deadline": "Cap negotiations at 50 days: balance hr timelines with cash runway and burn.",
+                "retention_plan": "Retention bonuses plus explicit reporting lines for VPs in core product.",
+            },
+            {
+                "counter_offer": "Signal 3.25x as the opening position with board ratified walk-down rights.",
+                "deadline": "Close or exit talks within 40 days; 7 month runway does not allow drift.",
+                "retention_plan": "Two-tier retention: cash now for chro-priority staff, earn-outs for the rest.",
+            },
+            {
+                "counter_offer": "Anchor at 3.1x, allow acquirer to meet near 2.7x with accelerated diligence.",
+                "deadline": "45-day triage: align cfo cash tests with cto and chro risk windows.",
+                "retention_plan": "Milestone-based retention: pay at signing, 90d, and close for key dev talent.",
+            },
+        ]
+        final = finals[v % len(finals)]
         return [
-            {"action_type": "delegate", "agent_id": "cto", "payload": "Assess IP leverage and technical moat."},
-            {"action_type": "delegate", "agent_id": "cfo", "payload": "Assess cash runway and valuation ceiling."},
-            {"action_type": "delegate", "agent_id": "chro", "payload": "Assess retention risk and people constraints."},
-            {"action_type": "log_reasoning", "payload": "CTO input supports a higher counter because the IP moat is meaningful."},
+            {"action_type": "delegate", "agent_id": "cto", "payload": d0},
+            {"action_type": "delegate", "agent_id": "cfo", "payload": d1},
+            {"action_type": "delegate", "agent_id": "chro", "payload": d2},
+            {"action_type": "log_reasoning", "payload": r_before_a[v % len(r_before_a)]},
             {
                 "action_type": "log_conflict",
                 "payload": json.dumps(
                     {
                         "id": "c1",
-                        "summary": "CTO valuation ambition conflicts with CFO runway and cash constraints.",
+                        "summary": c1_sums[v % len(c1_sums)],
                         "source_agents": ["cto", "cfo"],
                     }
                 ),
             },
-            {"action_type": "log_reasoning", "payload": "CFO input limits how long the negotiation can remain open."},
+            {"action_type": "log_reasoning", "payload": r_after_c1[v % len(r_after_c1)]},
             {
                 "action_type": "log_conflict",
                 "payload": json.dumps(
                     {
                         "id": "c2",
-                        "summary": "A slow process increases CHRO retention risk for key engineering talent.",
+                        "summary": c2_sums[v % len(c2_sums)],
                         "source_agents": ["chro", "cto"],
                     }
                 ),
@@ -420,17 +521,17 @@ def oracle_actions(task_id: str, variant: int = 0) -> List[Dict[str, Any]]:
                     {
                         "conflict_id": "c1",
                         "resolution_type": "bounded_counter",
-                        "text": "Ask above the CFO ceiling but set a fast deadline and walk-down logic.",
+                        "text": res_texts[v % len(res_texts)],
                     }
                 ),
             },
             {"action_type": "advance_phase", "payload": "analysis"},
-            {"action_type": "log_reasoning", "payload": "The 7 month runway makes delay costly and should shape the deadline."},
+            {"action_type": "log_reasoning", "payload": r_runway[v % len(r_runway)]},
             {"action_type": "advance_phase", "payload": "decision"},
-            {"action_type": "log_decision", "payload": "Proceed with a counter that acknowledges 7 month runway and cash limits."},
-            {"action_type": "log_reasoning", "payload": "Retention incentives are required before the market reads uncertainty."},
+            {"action_type": "log_decision", "payload": decision_line[v % len(decision_line)]},
+            {"action_type": "log_reasoning", "payload": r_retention[v % len(r_retention)]},
             {"action_type": "advance_phase", "payload": "execution"},
-            {"action_type": "log_reasoning", "payload": "Final recommendation integrates valuation, deadline, and retention plan."},
+            {"action_type": "log_reasoning", "payload": r_final[v % len(r_final)]},
             {"action_type": "finalize", "payload": json.dumps(final)},
         ]
     raise ValueError(f"unknown task_id: {task_id}")

scripts/prepare_sft_data.py

@@ -1,8 +1,7 @@
 """Convert verified CORP-ENV trajectories into chat-format SFT JSONL.
 
-Input should usually be `data/processed/e1_m1_clean.jsonl` from
-`scripts/verify_examples.py`. Each output row is compatible with TRL-style chat
-SFT datasets:
+Pass one or more processed JSONLs (e.g. `e1_m1_clean` + `h1_seed_clean`) from
+`scripts/verify_examples.py`. Each output row is TRL-style chat SFT data:
 
   {"task_id": "...", "example_id": "...", "messages": [...]}
 """
@@ -47,26 +46,64 @@ def convert_example(example: Dict[str, Any], min_pass_rate: float) -> Dict[str,
     }
 
 
+def _parse_input_paths(raw: List[str]) -> List[Path]:
+    """Expand comma-separated entries and return unique ordered paths."""
+    out: List[Path] = []
+    for part in raw:
+        for p in part.split(","):
+            p = p.strip()
+            if p:
+                out.append(Path(p))
+    return out
+
+
 def main() -> None:
     parser = argparse.ArgumentParser(description="Prepare chat SFT data from verified examples.")
-    parser.add_argument("--input", default="data/processed/e1_m1_clean.jsonl")
-    parser.add_argument("--output", default="data/sft/e1_m1_examples.jsonl")
+    default_inputs = (
+        "data/processed/e1_m1_clean.jsonl,data/processed/h1_seed_clean.jsonl"
+    )
+    parser.add_argument(
+        "--input",
+        dest="inputs",
+        action="append",
+        default=None,
+        metavar="PATH",
+        help=(
+            "Processed JSONL (repeat flag or use commas). "
+            f"Default: {default_inputs}"
+        ),
+    )
+    parser.add_argument("--output", default="data/sft/e1_m1_h1_examples.jsonl")
     parser.add_argument("--min-pass-rate", type=float, default=0.80)
     args = parser.parse_args()
+    raw_inputs = list(args.inputs) if args.inputs else [default_inputs]
+    input_paths = _parse_input_paths(raw_inputs)
 
     rows: List[Dict[str, Any]] = []
+    seen_ids: set[str] = set()
     skipped = 0
-    for example in read_jsonl(Path(args.input)):
-        try:
-            row = convert_example(example, args.min_pass_rate)
-        except Exception as exc:
-            skipped += 1
-            print(f"skip {example.get('example_id', 'unknown')}: {exc}")
-            continue
-        if row is None:
-            skipped += 1
+    for path in input_paths:
+        if not path.is_file():
+            print(f"warning: input missing, skip: {path}", file=sys.stderr)
             continue
-        rows.append(row)
+        for example in read_jsonl(path):
+            eid = str(example.get("example_id") or example.get("id") or "")
+            if eid and eid in seen_ids:
+                skipped += 1
+                continue
+            try:
+                row = convert_example(example, args.min_pass_rate)
+            except Exception as exc:
+                skipped += 1
+                print(f"skip {example.get('example_id', 'unknown')}: {exc}")
+                continue
+            if row is None:
+                skipped += 1
+                continue
+            rows.append(row)
+            eid2 = str(row.get("example_id") or "unknown")
+            if eid2 and eid2 != "unknown":
+                seen_ids.add(eid2)
 
     write_jsonl(Path(args.output), rows)
     print(f"Wrote {len(rows)} SFT conversations to {args.output}; skipped {skipped}.")

tests/test_pipeline.py

@@ -32,6 +32,17 @@ class PipelineSmokeTests(unittest.TestCase):
         self.assertEqual(result.status, "clean")
         self.assertGreaterEqual(result.verifier_pass_rate, 5 / 6)
 
+    def test_oracle_h1_variants_replay_cleanly(self) -> None:
+        for v in (0, 5, 11):
+            with self.subTest(variant=v):
+                result = replay_actions(
+                    example_id=f"test-h1-{v}",
+                    task_id="h1_acquisition_defence",
+                    actions=oracle_actions("h1_acquisition_defence", v),
+                )
+                self.assertEqual(result.status, "clean")
+                self.assertGreaterEqual(result.verifier_pass_rate, 7 / 11)
+
     def test_invalid_trajectory_is_rejected(self) -> None:
         result = replay_actions(
             example_id="bad",

training/train_grpo.py

@@ -11,9 +11,9 @@ Run on Colab, Lightning AI H100, or another GPU machine after SFT:
   python training/train_grpo.py \
     --model Qwen/Qwen2.5-7B-Instruct \
     --adapter outputs/sft_Qwen2.5-7B-Instruct \
-    --examples data/processed/e1_m1_clean.jsonl \
+    --examples data/processed/e1_m1_clean.jsonl,data/processed/h1_seed_clean.jsonl \
     --output outputs/grpo_Qwen2.5-7B-Instruct \
-    --max-steps 50 \
+    --max-steps 150 \
     --push-to-hub your-org/corp-gym-grpo-qwen2.5-7b
 
 The reward function recreates the environment state from a verified action prefix,
@@ -30,6 +30,11 @@ Optional speedups (Unsloth picks these up automatically when importable):
 - **xFormers**: already used as a fallback when FA2 is missing; still slower than FA2.
 - **Qwen3.x “linear attention” fast path**: only for those architectures; not used for
   Qwen2.5 GRPO. See Unsloth logs if you train Qwen3.5+.
+
+On a Linux H100 (e.g. Lightning), try **larger** `--batch-size` if memory allows, and
+`--dataloader-num-workers 2`–`4` (Windows often keeps 0) so the `CorpEnvironment` rollouts
+are not the only bottleneck. Use `--max-steps` (default 150) and `--generations` to trade
+quality vs wall-clock.
 """
 
 from __future__ import annotations
@@ -72,12 +77,17 @@ def prompt_for_prefix(task_id: str, prefix_actions: List[Dict[str, Any]]) -> Lis
     return messages
 
 
+def _examples_paths(examples_path: str) -> List[Path]:
+    return [Path(p.strip()) for p in examples_path.split(",") if p.strip()]
+
+
 def build_prompt_dataset(examples_path: str, tasks: List[str], repeats: int) -> List[Dict[str, Any]]:
+    """Load examples from one or more JSONLs (comma-separated). Falls back to oracle prefixes if all missing/empty."""
     rows: List[Dict[str, Any]] = []
-    path = Path(examples_path)
-    if path.exists():
-        examples = list(read_jsonl(path))
-        for example in examples:
+    for path in _examples_paths(examples_path):
+        if not path.exists():
+            continue
+        for example in read_jsonl(path):
             if example.get("status") and example.get("status") != "clean":
                 continue
             task_id = str(example.get("task_id") or "")
@@ -157,9 +167,16 @@ def main() -> None:
     parser = argparse.ArgumentParser(description="Train CORP-ENV GRPO adapter.")
     parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct")
     parser.add_argument("--adapter", default="outputs/sft_Qwen2.5-7B-Instruct")
-    parser.add_argument("--examples", default="data/processed/e1_m1_clean.jsonl")
+    parser.add_argument(
+        "--examples",
+        default="data/processed/e1_m1_clean.jsonl,data/processed/h1_seed_clean.jsonl",
+        help="One JSONL or comma-separated list of verified (clean) trajectory files.",
+    )
     parser.add_argument("--output", default="outputs/grpo_Qwen2.5-7B-Instruct")
-    parser.add_argument("--tasks", default="e1_launch_readiness,m1_budget_reallocation")
+    parser.add_argument(
+        "--tasks",
+        default="e1_launch_readiness,m1_budget_reallocation,h1_acquisition_defence",
+    )
     parser.add_argument(
         "--repeats",
         type=int,
@@ -180,8 +197,8 @@ def main() -> None:
     parser.add_argument(
         "--max-steps",
         type=int,
-        default=50,
-        help="Optimizer steps (not env episodes). Lower finishes sooner; raise for stronger fit.",
+        default=150,
+        help="Optimizer steps (not env episodes). Default 150 for a fuller RL fit on H100-style runs.",
     )
     parser.add_argument(
         "--save-steps",
@@ -190,6 +207,18 @@ def main() -> None:
         help="Checkpoint every N steps. Default: min(25, max(5, max_steps//2)) so short runs still save.",
     )
     parser.add_argument("--optim", default="adamw_8bit")
+    parser.add_argument(
+        "--dataloader-num-workers",
+        type=int,
+        default=0,
+        help="DataLoader workers (0 is safest on Windows; try 2–4 on Linux H100 if CPU allows).",
+    )
+    parser.add_argument(
+        "--dataloader-prefetch-factor",
+        type=int,
+        default=None,
+        help="When dataloader_num_workers>0, optional prefetch depth (e.g. 2).",
+    )
     parser.add_argument("--push-to-hub", default="")
     args = parser.parse_args()
 
@@ -242,24 +271,29 @@ def main() -> None:
             random_state=3407,
         )
 
-    config = GRPOConfig(
-        output_dir=args.output,
-        learning_rate=args.lr,
-        per_device_train_batch_size=args.batch_size,
-        gradient_accumulation_steps=args.grad_accum,
-        num_generations=args.generations,
-        max_prompt_length=args.max_prompt_length,
-        max_completion_length=args.max_completion_length,
-        max_steps=args.max_steps,
-        logging_steps=5,
-        save_steps=save_steps,
-        save_total_limit=3,
-        optim=args.optim,
-        bf16=True,
-        report_to="none",
-        push_to_hub=bool(args.push_to_hub),
-        hub_model_id=args.push_to_hub or None,
-    )
+    _gc_kwargs: Dict[str, Any] = {
+        "output_dir": args.output,
+        "learning_rate": args.lr,
+        "per_device_train_batch_size": args.batch_size,
+        "gradient_accumulation_steps": args.grad_accum,
+        "num_generations": args.generations,
+        "max_prompt_length": args.max_prompt_length,
+        "max_completion_length": args.max_completion_length,
+        "max_steps": args.max_steps,
+        "logging_steps": 5,
+        "save_steps": save_steps,
+        "save_total_limit": 3,
+        "optim": args.optim,
+        "bf16": True,
+        "report_to": "none",
+        "push_to_hub": bool(args.push_to_hub),
+        "hub_model_id": args.push_to_hub or None,
+    }
+    if args.dataloader_num_workers:
+        _gc_kwargs["dataloader_num_workers"] = args.dataloader_num_workers
+    if args.dataloader_prefetch_factor is not None and args.dataloader_num_workers:
+        _gc_kwargs["dataloader_prefetch_factor"] = args.dataloader_prefetch_factor
+    config = GRPOConfig(**_gc_kwargs)
     trainer = GRPOTrainer(
         model=model,
         processing_class=tokenizer,

training/train_sft.py

@@ -4,12 +4,14 @@ This is the hackathon SFT training script. It uses:
 
 - `unsloth.FastLanguageModel` for efficient 4-bit LoRA/QLoRA loading.
 - `trl.SFTTrainer` / `trl.SFTConfig` for supervised fine-tuning.
+- `messages`-format JSONL with TRL 0.2x conversational SFT (optional `--assistant-only` if the
+  model chat template supports assistant token masks; Qwen2.5 Instruct defaults to off).
 
 Run on Colab, Lightning AI H100, or another GPU machine:
 
   python training/train_sft.py \
     --model Qwen/Qwen2.5-7B-Instruct \
-    --data data/sft/e1_m1_examples.jsonl \
+    --data data/sft/e1_m1_h1_examples.jsonl \
     --output outputs/sft_adapter \
     --max-steps 30 \
     --push-to-hub your-org/corp-env-sft-adapter
@@ -18,9 +20,37 @@ Run on Colab, Lightning AI H100, or another GPU machine:
 from __future__ import annotations
 
 import argparse
+import inspect
 import json
+from dataclasses import fields
 from pathlib import Path
-from typing import Dict, List
+from typing import Any, Dict, List
+
+import torch  # imported before unsloth in main for dtype hooks
+
+
+def _sft_config_field_names() -> set[str]:
+    from trl import SFTConfig
+
+    if hasattr(SFTConfig, "__dataclass_fields__"):
+        return set(SFTConfig.__dataclass_fields__.keys())
+    return {f.name for f in fields(SFTConfig)}
+
+
+def load_conversation_rows(path: Path) -> List[Dict[str, Any]]:
+    """TRL conversational format: each row has a `messages` list."""
+    rows: List[Dict[str, Any]] = []
+    with path.open("r", encoding="utf-8") as f:
+        for line in f:
+            if not line.strip():
+                continue
+            obj = json.loads(line)
+            if "messages" not in obj:
+                raise SystemExit(
+                    f"{path}: SFT example missing 'messages' (use --legacy-text for old format)."
+                )
+            rows.append({"messages": obj["messages"]})
+    return rows
 
 
 def load_text_rows(path: Path, tokenizer: object) -> List[Dict[str, str]]:
@@ -40,10 +70,81 @@ def load_text_rows(path: Path, tokenizer: object) -> List[Dict[str, str]]:
     return rows
 
 
+def _build_sft_config(
+    allowed: set[str],
+    output_dir: str,
+    max_seq: int,
+    args: argparse.Namespace,
+) -> Any:
+    from trl import SFTConfig
+
+    candidate: Dict[str, Any] = {
+        "output_dir": output_dir,
+        "max_length": max_seq,
+        "per_device_train_batch_size": args.batch_size,
+        "gradient_accumulation_steps": args.grad_accum,
+        "num_train_epochs": args.epochs,
+        "learning_rate": args.lr,
+        "warmup_ratio": 0.05,
+        "lr_scheduler_type": "cosine",
+        "logging_steps": 5,
+        "save_steps": args.save_steps,
+        "save_total_limit": 3,
+        "max_steps": args.max_steps,
+        "optim": args.optim,
+        "bf16": (not args.fp16) and torch.cuda.is_available(),
+        "fp16": bool(args.fp16) and torch.cuda.is_available(),
+        "packing": args.packing,
+        "report_to": "none",
+        "push_to_hub": bool(args.push_to_hub),
+        "hub_model_id": args.push_to_hub or None,
+        "assistant_only_loss": args.assistant_only,
+    }
+    if args.dataset_num_proc is not None and "dataset_num_proc" in allowed:
+        candidate["dataset_num_proc"] = args.dataset_num_proc
+    if args.dataloader_num_workers and "dataloader_num_workers" in allowed:
+        candidate["dataloader_num_workers"] = args.dataloader_num_workers
+    if args.packing and "padding_free" in allowed and args.padding_free:
+        candidate["padding_free"] = True
+    if args.legacy_text and "dataset_text_field" in allowed:
+        candidate["dataset_text_field"] = "text"
+
+    kwargs = {k: v for k, v in candidate.items() if k in allowed}
+    if "output_dir" not in kwargs:
+        kwargs["output_dir"] = output_dir
+    return SFTConfig(**kwargs)
+
+
+def _build_trainer(
+    model: object,
+    tokenizer: object,
+    config: object,
+    dataset: object,
+) -> Any:
+    from trl import SFTTrainer
+
+    sig = inspect.signature(SFTTrainer.__init__)
+    if "processing_class" in sig.parameters:
+        return SFTTrainer(
+            model=model,
+            args=config,
+            train_dataset=dataset,
+            processing_class=tokenizer,
+        )
+    if "tokenizer" in sig.parameters:
+        return SFTTrainer(
+            model=model,
+            args=config,
+            train_dataset=dataset,
+            tokenizer=tokenizer,
+        )
+    raise SystemExit("SFTTrainer: expected processing_class or tokenizer in __init__.")
+
+
 def main() -> None:
     parser = argparse.ArgumentParser(description="Train CORP-ENV SFT LoRA adapter.")
     parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct")
-    parser.add_argument("--data", default="data/sft/e1_m1_examples.jsonl")
+    parser.add_argument("--data", default="data/sft/e1_m1_h1_examples.jsonl")
     parser.add_argument("--output", default="outputs/sft_adapter")
     parser.add_argument("--max-seq-length", type=int, default=8192)
     parser.add_argument("--epochs", type=float, default=2.0)
@@ -54,19 +155,61 @@ def main() -> None:
     parser.add_argument("--save-steps", type=int, default=50)
     parser.add_argument("--optim", default="adamw_8bit")
     parser.add_argument("--push-to-hub", default="")
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Use fp16 instead of bf16 (bf16 is default on CUDA when available).",
+    )
+    parser.add_argument(
+        "--legacy-text",
+        action="store_true",
+        help="Pre-tokenize to a single 'text' column (older path). Default: TRL messages format.",
+    )
+    parser.add_argument(
+        "--assistant-only",
+        action="store_true",
+        help=(
+            "Loss only on assistant tokens (requires tokenizer chat template with assistant masks). "
+            "Qwen2.5 usually needs this off; use with templates that support {% generation %}."
+        ),
+    )
+    parser.add_argument(
+        "--packing",
+        action="store_true",
+        help="Sequence packing (faster; avoid for very long or uneven traces unless you know the tradeoffs).",
+    )
+    parser.add_argument(
+        "--padding-free",
+        action="store_true",
+        help="With packing+bfd, TRL can use padding-free; requires FlashAttention-friendly setup.",
+    )
+    parser.add_argument(
+        "--dataset-num-proc",
+        type=int,
+        default=4,
+        help="Multiprocess dataset preprocessing; set 0 to disable where supported (default 4 for H100).",
+    )
+    parser.add_argument(
+        "--dataloader-num-workers",
+        type=int,
+        default=2,
+        help="DataLoader workers (default 2; increase on H100 if CPU allows).",
+    )
     args = parser.parse_args()
 
     try:
-        # Unsloth must load before trl/transformers/peft for full optimizations.
         from unsloth import FastLanguageModel
         from datasets import Dataset
-        from trl import SFTConfig, SFTTrainer
     except ImportError as exc:
         raise SystemExit(
             "SFT training requires datasets, trl, and unsloth. On Lightning AI, install with:\n"
-            "  pip install -U unsloth trl datasets accelerate peft bitsandbytes transformers"
+            "  pip install -e \".[training]\""
         ) from exc
 
+    allowed = _sft_config_field_names()
+    if args.dataset_num_proc == 0 and "dataset_num_proc" in allowed:
+        args = argparse.Namespace(**{**vars(args), "dataset_num_proc": None})
+
     model, tokenizer = FastLanguageModel.from_pretrained(
         model_name=args.model,
         max_seq_length=args.max_seq_length,
@@ -97,34 +240,16 @@ def main() -> None:
         random_state=3407,
     )
 
-    dataset = Dataset.from_list(load_text_rows(Path(args.data), tokenizer))
-    config = SFTConfig(
-        output_dir=args.output,
-        dataset_text_field="text",
-        max_length=args.max_seq_length,
-        per_device_train_batch_size=args.batch_size,
-        gradient_accumulation_steps=args.grad_accum,
-        num_train_epochs=args.epochs,
-        learning_rate=args.lr,
-        warmup_ratio=0.05,
-        lr_scheduler_type="cosine",
-        logging_steps=5,
-        save_steps=args.save_steps,
-        save_total_limit=3,
-        max_steps=args.max_steps,
-        optim=args.optim,
-        bf16=True,
-        packing=False,
-        report_to="none",
-        push_to_hub=bool(args.push_to_hub),
-        hub_model_id=args.push_to_hub or None,
-    )
-    trainer = SFTTrainer(
-        model=model,
-        tokenizer=tokenizer,
-        train_dataset=dataset,
-        args=config,
-    )
+    if args.legacy_text:
+        rows = load_text_rows(Path(args.data), tokenizer)
+    else:
+        rows = load_conversation_rows(Path(args.data))
+    if not rows:
+        raise SystemExit(f"No training rows in {args.data!r}.")
+
+    dataset = Dataset.from_list(rows)
+    config = _build_sft_config(allowed, args.output, args.max_seq_length, args)
+    trainer = _build_trainer(model, tokenizer, config, dataset)
     trainer.train()
     trainer.save_model(args.output)
     tokenizer.save_pretrained(args.output)