Spaces:
Sleeping
Sleeping
Commit History
-
feat(phase1): OpenEnv scaffold + R1/R2 rewards — PHASE 1 GATE PASS
- environment/actions.py: ActionType enum + ArbitratorAction model - environment/observations.py: RewardComponents (weighted, normalised), DebateRound, Observation - environment/episode_state.py: mutable episode state dataclass - environment/env.py: ViralScriptEnv — Gymnasium-compatible reset/step/state, difficulty tiers, anti-gaming wired - rewards/r1_hook_strength.py: 5-check rule-based hook scorer (promise, curiosity, specificity, front-load, anti-filler) - rewards/r2_coherence.py: sentence-transformers cosine similarity with 4-range score mapping + embedding cache - rewards/reward_aggregator.py: catastrophic-drop (>0.2) + action-diversity anti-gaming rules - agents/rewriter.py: RewriterAgent wrapping LLMBackend with unified diff output - scripts/run_dummy_episode.py: demo runner with rich output and gate check - tests/test_rewards.py: 10 tests — R1 (5), R2 (2), aggregator (3) - tests/test_environment.py: 6 tests — all LLM calls mocked via golden fixtures Co-Authored-By: Claude Son…41ea373