reality-feedback-createlisted

# Reality-feedback auto-create skill Layer 4 of the layered eval architecture. Auto-creates ledger entries for high-stakes outputs so their outcome can be graded later. The discipline is: at output time, capture the prediction. At outcome time (days/weeks later), grade actual vs predicted. Over months, the deltas feed a Layer 1 v2 classifier's training set. ## Example output (ledger entry shape) ```json { "id": "rfb-2026-05-11-001", "ts_created": "2026-05-11T05:40:00+00:00", "session_id": "s_2026-05-11-build", "stake_class": "high", "stake_signals": ["draft", "outbound", "high-value-proposal"], "predicted_surface": "counter-offer should test a higher retainer + shorter term", "outcome_window_days": 30, "outcome_due": "2026-06-10", "outcome": null, "critic_metadata": {"score": 8.4, "passes": 3, "concerns": ["could be premature pattern-lock at n=2"]} } ``` ## Conversation context (prior) **Auto-fired** by the Stop hook chain. runs AFTER classifier_dispatch + critic_dispatch + (optionally) save_handler. The prior conversation is the just-completed turn whose output triggered the high-stake classification. The script walks `logs/session.jsonl` back to find the most-recent `stake_classified` event with `class=high` + the matching tool-output that produced it. ## Output format Returns ONE row appended to `data/reality_feedback.sqlite` table `outcomes`: | column | type | description | |---|---|---| | id | TEXT PK | `rfb-YYYY-MM-DD-NNN` | | ts_created | IS