Commit History
-
data: regenerate level3.jsonl — 100 questions with adversarial pressure messages
Jayant-KernelUnverified signature43ba213 -
data: add level3.jsonl — 100 questions with adversarial pressure messages
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signatureCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>83e9993 -
fix: persist initial context (distractors/pressure) across all episode turns
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signatureCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>725414c -
feat: extend environment with level=3, resistance reward ±0.2
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signatureAdds Level 3 adversarial pressure support: level3 dataset loading, pressure_messages context injection, pressure_shown state tracking, and a RESISTANCE_BONUS=0.2 reward component in the terminal step. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>27deeb6 -
feat: add pressure_shown field to DeceitState
Jayant-KernelUnverified signatureed81b09 -
Phase 4 complete: level2 dataset, distractor env, 90 tests passing
Jayant-KernelUnverified signature46011b7 -
feat: extend reset() to support level=2 with distractor context
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signatureCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f2049f5 -
feat: add 429 retry wrapper to grader semantic check
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signatureCo-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b44d7b0 -
Upload folder using huggingface_hub
Verified signature61af0e3
-
Phase 2.5: multi-turn episodes, bug fixes, dataset cleanup
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signatureCritical fixes: - Multi-turn episodes (max_turns=3): is_final flag on DeceitAction, step penalty -0.05 per thinking turn, prior reasoning fed back as context - Grader cache: defaults to /tmp/deceit_grader_cache.json (HF Spaces safe), overridable via DECEIT_GRADER_CACHE env var - Dataset: replaced 6 ambiguous questions (q051,q058,q066,q072,q091,q100) Other improvements: - DeceitState adds prior_reasoning: list[str] and max_turns: int - DeceitAction adds is_final: bool = False - __init__.py: package-level exports for DeceitEnvironment, models, Grader - app.py: factory pattern limitation documented - test_models.py: covers inherited OpenEnv fields, is_final, extra field rejection - test_environment.py: multi-turn trajectory, forced termination, step penalties - 72 tests passing (up from 56) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9737348 -
Phase 2 complete: Level 1 env runs locally, tests green, 100-question dataset
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signature- models.py: upgraded to real OpenEnv base classes (Action/Observation/State) - grader.py: exact-match + GPT-4o-mini semantic fallback with disk cache - environment.py: DeceitEnvironment inheriting openenv.core.env_server.Environment - reset/step/state implementing full episode loop - compute_reward() pure function (correctness + calibration signals) - app.py: create_fastapi_app wrapper exposing /reset /step /state - level1.jsonl: 100 hand-curated QA pairs across 5 categories - 56 tests passing: models, rewards, grader (mocked), environment (mocked) - Smoke test: +1.3 correct+confident, -1.3 wrong+confident, 0.0 abstain OpenEnv API note: step() returns Observation with reward embedded (not tuple). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f577d1f -
Phase 1 complete: schemas, reward design, project scaffold
Jayant-Kernel +1: Claude Sonnet 4.6Unverified signature- Pydantic v2 schemas for DeceitObservation, DeceitAction, DeceitState - REWARD_DESIGN.md with correctness + calibration signals, worked examples, gaming analysis - Full folder structure with Phase 2-6 stubs - pytest tests for all three schemas (validation, defaults, JSON roundtrip) - pyproject.toml with openenv-core[core]>=0.2.1 dependency pinned for Phase 2 - Design doc in docs/superpowers/specs/ Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>139d3d1