SE_STATE.md

SE State (Reference)

This short reference documents the canonical artifacts and files used to observe SE runs and their verification status.

Primary run artifacts

File	Contents
checklist_draft.json	Draft checklist; {"items": [...], "status": "draft"}. Written by main.run_self_host() / _dsl_items() before enrichment.
checklist.json	Promoted checklist (same structure); written when draft items are accepted.
checklist_quality.json	Quality evaluation: per-item quality_status (VALID/INVALID), reasons, low-signal flags, totals.
requirement_completeness.json	Per-requirement coverage: state (COMPLETE/PARTIAL/UNBOUND), complete_pct.
spec_coverage.json	Per-spec-pointer coverage: coverage_pct, uncovered_units.
checklist_sufficiency.json	Sufficiency verdict: sufficient, complete_pct, grade.
manifest.json	Run-level manifest: item count, quality summary, readiness grade.
gap_report.json	Gap taxonomy written by learning/gap_analyzer.py during learning trials.

Recommended checks for CI

• Confirm checklist_draft.json exists and items list is non-empty.
• Confirm checklist_quality.json shows invalid_items == 0 for release-ready specs.
• Confirm requirement_completeness.json complete_pct >= 0.98 for tier-2+ specs.
• For deterministic runs: verify manifest.json fingerprint matches baseline.

Self-learning artifacts

The learning trial writes additional artifacts under learning_artifacts/iter_N/{spec_stem}/:

File	Contents
learning_artifacts/iter_N/iteration_summary.json	corpus_score, specs_met, converged, per-spec scores.
learning_artifacts/iter_N/improvement_prompts.md	LLM-ready gap prompts with live source snippets.
learning_artifacts/history.json	Iteration-over-iteration corpus_score trend.

Current status (2026-03-03)

• corpus_score = 1.0000, specs_met = 32/32, converged = True
• Fuzz suite: 2000/2000 random specs pass at quality ≥ 80% and completeness ≥ 70%