CERTIFICATION.mdis the one authored file here — the forefst soundness certification (v3.5.0, 2026-07-01): the verdict, what was delivered, the verification gates, and the audit trail. Everything else in this directory is machine-regenerated evidence.
Machine-regenerated verification outputs — not authored documentation. They are overwritten on every gate run; treat them as the latest evidence snapshot, not a stable reference.
| File | Produced by | What it shows |
|---|---|---|
report_static_claims.txt |
verify_docs_static.py |
Static-claim checks (function/symbol counts, version gating) — PASS/WARN/FAIL |
report_tool_execution.txt |
verify_docs_tools.py |
Tool execution log (commands, exit codes, timing) over the image corpus |
report_tool_claims.txt |
verify_docs_tools.py |
Tool-verified claim results — PASS/FAIL/SKIP |
report_summary.txt |
manual roll-up | Pre-publication summary |
report_static_verification.txt |
verify_static.py |
Static-evidence pass over the reference-table NOT_TESTED entries (decompilation + function-catalog search) — PASS/WARN/FAIL |
report_oid_0x520_*.txt |
ad-hoc probe (investigate_oid_0x520.py) |
Historical OID 0x520 (Change Journal / FS Metadata directory) investigation |
Regenerate (from the repo root): python3 analysis/reports/verify_docs_static.py and python3 analysis/reports/verify_docs_tools.py.
Regeneration runs only against the private disk corpus and the Ghidra / decompiled exports, which are not bundled in this repository; point the scripts at a local copy via
REFS_DISKS/REFS_CORPUS/REFS_GHIDRA/REFS_DECOMPILED/REFS_FUNC_CATALOG.The ~29,761-row function catalog is not bundled.
verify_docs_static.pyresolves it in order:REFS_FUNC_CATALOG→forefst/analysis/function_catalog.csv(drop a copy here) →<corpus>/forclaude/intelligence/function_catalog.csv(the workspace copy). If none is found it exits non-zero with a message naming the env override — it no longer aborts with a traceback.
A third gate, verify_tool_tables.py (in this directory) — reference-table consistency, asserts the tools' embedded opcode/schema/
OID/flag name maps against the master + decompilation), is stdout-only and writes no report file here; run it
alongside the two above (--calibrate proves it fires on planted errors).
Three more scripts in this directory verify reference-table claims directly: verify_static.py produces report_static_verification.txt (above); verify_usn_claims.py (USN / Change-Journal claims) and verify_not_tested.py (the remaining both-NOT_TESTED entries) are stdout-only and write no report file here.
Where the reports above are machine-regenerated gate output, audit/ holds the
authored audit deliverables — the record of how the published knowledge and tools were validated
against ground truth (raw-disk bytes + decompilation), adversarially and with calibrated instruments.
It covers four audits: references & central findings, the analysis tree, the documentation, and
the tools. Full account in audit/README.md; the key contents:
| Path | What it is |
|---|---|
AUDIT_COMPLETE.md |
Headline result — 409/409 reference-table claims audited, 0 byte-layout failures |
proof_index.csv |
Master crosswalk: each claim → class, verdict, static proof, disk validation, dossier |
dossiers/<ref_id>.md |
Per-claim dossier — canonical claim, verdict, static + disk proof (one per claim) |
proofs/static/ |
Exported decompiled functions backing the static/decompilation claims |
proofs/validation/<ref_id>.csv |
Per-claim disk-corpus matrix: one row per applicable image (path,basename,group,result,value). Header-only when a claim is proven by citation/decompilation rather than disk measurement — so ~285/411 are header-only by design, not missing data |
audit_harness.py + specs.jsonl |
Regenerate every matrix/dossier/index from the authored specs (python3 audit/audit_harness.py) |
tool_audit.md, analysis_consistency_audit.md |
The tool-correctness audit and analysis-tree consistency audit write-ups |
See also: repo README