← ClaudeAtlas

rnd-calibrationlisted

Use when recording verdict data, detecting false verdicts, or injecting calibration stats into verifier prompts — enables the framework to learn from verification mistakes over time
oleksify/rnd-framework · ★ 0 · AI & Automation · score 75
Install: claude install-skill oleksify/rnd-framework
# R&D Calibration ## Overview Track verification verdicts over time so the framework can detect systematic bias — verifiers that pass broken work or fail correct work. Calibration data lives in a simple JSONL file at the project base level. **Core principle:** A verifier that issues false verdicts is worse than no verifier. Calibration surfaces patterns before they become systemic. ## When to Use - After any task completes a full build-verify cycle (automated via orchestrator) - When a manual ground-truth correction is needed (`/rnd-framework:rnd-calibrate`) - When the orchestrator constructs verifier prompts (inject calibration summary) - When diagnosing a pipeline that keeps cycling without converging ## JSONL Record Schema Each completed task appends one record to `calibration.jsonl`: ```json { "task_id": "M1.T03.example-task", "sessionId": "20260316-154145-1227", "verdict": "PASS", "criterionResults": [ { "criterion": "File exists at path", "result": "PASS" }, { "criterion": "YAML frontmatter valid", "result": "PASS" } ], "iterationCount": 2, "timestamp": "2026-03-16T15:41:45Z", "falseVerdictFlag": null } ``` **Fields:** | Field | Type | Notes | |-------|------|-------| | `task_id` | string | Task identifier (e.g. `"M1.T03.example-task"`) | | `sessionId` | string | Session that produced this verdict | | `verdict` | string | `"PASS"`, `"FAIL"`, `"NEEDS_ITERATION"`, or `"PASS_QUALITY_NEEDS_ITERATION"` | | `criticality` | string | `"LOW"`, `