-
fix: clamp ALL rewards/scores to strict (0.01, 0.99) — every output path
-
feat: 3 tasks with programmatic graders + OPENAI_API_KEY support
Required by hackathon spec:
- Task 1: single_shipment (Easy) — grader: 1.0/0.5/0.0
- Task 2: disruption_reroute (Medium) — grader: delivery*0.7 + on_time*0.3
- Task 3: crisis_management (Hard) — grader: 0.50*delivery - 0.30*loss - 0.20*cost
- inference.py reads OPENAI_API_KEY, runs all 3 tasks by name
- 19 tests passing -
feat: SupplyChainEnv — global supply chain disruption RL environment
ragavrida +1: Claude Opus 4.6 (1M context)
