{"data":{"kind":"file","path":"README.md","version_id":"g71ju56pjyrmj271s54xe8dy","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1011,"modified_at":"2026-02-18T12:54:53.337000","content_hash":"e1d302b36eeb42c4deed0013217743b2b95387026fd1c5eedef6bdb30ff8eaf4"},"entries":[],"content":"# OODA Decision Environment\n\nVerifiers-compatible environment for training autonomous agents on OODA loop decision-making.\n\n## What it does\n\nPresents observations + orientation context and expects the model to:\n1. Choose the correct action type (`execute`, `research`, `escalate`, `delegate`, `schedule`, `noop`)\n2. Assign a calibrated confidence score (0.0–1.0)\n3. Explain its reasoning\n\n## Reward function\n\n- 60% action correctness (binary: right or wrong, with partial credit for related actions)\n- 25% confidence calibration (gated: only when action is correct)\n- 15% format compliance (structured `<decision>` format)\n\n## 33 curated scenarios\n\nCovers Flow/Rise agent operations and Seq desktop context with deterministic split support for production training.\n\n- `train`: policy updates\n- `val`: online checkpoint selection\n- `test`: holdout generalization\n\n## Usage\n\n```bash\nprime env install nikiv/ooda-decision\n```\n\n```toml\n[[env]]\nid = \"nikiv/ooda-decision\"\nargs = { split = \"train\", seed = 42 }\n```\n","encoding":"utf-8","truncated":false,"total_bytes":1011},"status":null}