REG-D10 / AUD-06 — FIXED   (retracted 10k / p99 1,189.89 ms pair presented as measured)
docs/ + scripts/ + tests/ · evidence captured 2026-09-21 on this host

FINDING (AUDIT_REPORT_v5.0.1_PREP.md §AF-007, AF-009, AF-010, AF-011)
  The claims register retracts the v3.1.0 10,000-record / p99 1,189.89 ms
  backpressure pair (`UC-018`: the committed artifact contains 2,500 records at
  p99 836.3514210795984 ms), but buyer-facing and canonical documents still
  presented it as the retained measurement. No artifact in this tree produces
  the pair.

SWEEP (fourteen sites; the audit named twelve)
  replaced with the artifact-backed figures (read from the JSON, not restated):
    2026-08-20 run  evidence/execution_2026-08-20/backpressure_stall_report.json
                    2,500 offered / 2,500 durable, elapsed 6.630050110979937 s,
                    p50 167.29020897764713, p95 504.7037289477885,
                    p99 836.3514210795984, max 2290.622486965731 ms, 2,501 fsync calls
    2026-09-16 runs evidence/execution_2026-09-16/backpressure_stall_report_run{1,2,3}.json
                    2,500 offered / 2,500 durable, p99 52.317 / 51.875 / 47.531 ms
  site list (all now carry the retraction marker or the in-tree figures):
    DEPLOYMENT_GUIDE.md:138 · docs/BENCHMARKS.md:36,:37,:105,:107 ·
    docs/FAQ_PROCUREMENT.md:64 · docs/FAQ_TECHNICAL.md:33,:118 ·
    docs/PLATFORM_OPERATOR_GUIDE.md:163 (found by this sweep, not in the audit list) ·
    docs/PRODUCT_BRIEF_US.md:43 · docs/PROSPECTUS.md:58 · docs/ROADMAP.md:80 ·
    docs/benchmarks/BENCHMARK_METHOD.md:110 · docs/benchmarks/BENCHMARK_RESULTS.md:14,:15,:63-67 ·
    docs/benchmarks/README.md:43 · docs/operations/BACKPRESSURE_RUNBOOK.md:53 ·
    docs/performance/SCALING_GUIDE.md:38

GATE (`scripts/verify_claims.py` check 9, CI-wired at .github/workflows/ci.yml:110)
  A retracted token (`1,189.89` / `1189.89` / `10,000 offered requests` /
  `10,000 durable`) must carry its retraction on the same line or — for wrapped
  prose — the same paragraph; markdown table rows are checked individually so
  one row's retraction cannot cover the next row's citation. The audit report is
  exempt: it quotes the defective text as its evidence.
  $ .venv/bin/python scripts/verify_claims.py --root .
  verify_claims: PASS (102 claims, 0 findings)
  Negative control (live, then reverted — the probe file is not in git status):
  $ printf 'Live negative control: the run recorded p99 1,189.89 ms.\n' > docs/_tmp_gate_probe.md
  $ .venv/bin/python scripts/verify_claims.py --root .
  verify_claims: FAIL (102 claims, 1 findings)
    [retracted-figure-cited] UC-018
        docs/_tmp_gate_probe.md:1 cites '1,189.89' without its retraction (UC-018)
        on the same line or paragraph
  exit=1
  $ rm docs/_tmp_gate_probe.md && .venv/bin/python scripts/verify_claims.py --root .
  verify_claims: PASS (102 claims, 0 findings)

TEST (`tests/test_retracted_figure_gate.py`, 7 tests)
  $ .venv/bin/python -m pytest tests/test_retracted_figure_gate.py -q
  7 passed in 0.21s
  Covers: unmarked latency citation, unmarked 10,000-record claim, marked
  retraction allowed, wrapped paragraph as one unit, table rows individual,
  exempt audit report, artifact-backed figures not flagged.

BATTERY ON THIS COMMIT'S CONTENT (recorded before the final amend; the amend changed only this file)
$ AEGIS_SECURITY_ENFORCEMENT_MODE=development HERMES_SANDBOX=true .venv/bin/python -m pytest tests/ -n auto -q
6896 passed, 118 skipped in 85.33s (0:01:25)   [recorded run, exit 0]
  (Runs of this command started from the assistant's Python execution kernel
   record "6788 passed, 118 skipped, 5 errors": that kernel exports
   PYTHONPATH=/home/luna/.hermes/cache/scratch/hermes_kernel_<id>:/home/luna/.hermes/hermes-agent,
   which leaks into the pytest subprocess and makes five modules fail to import
   (tests/test_documentation_verifier.py, test_iq_oq.py, test_threat_lab.py,
   test_visualizer_app.py, test_wal_commit_span.py — 108 tests never collected).
   Measured as a controlled pair: same command, same commit, PYTHONPATH unset ->
   6896 passed, 118 skipped, exit 0; with the kernel PYTHONPATH injected ->
   6788 passed, 118 skipped, 5 errors. The clean figure is the shell run, which
   is what the battery convention uses. The contamination is in the harness,
   not in the repository.)
$ .venv/bin/bandit -r aegis/ aegis_server/ -c pyproject.toml -lll   -> Medium: 0, High: 0
$ .venv/bin/python scripts/verify_import_reachability.py            -> PASS (no undeclared orphans, no stale roadmap entries)
$ .venv/bin/python scripts/verify_release_contract.py --root .      -> 14 synchronized anchors at 5.0.0
$ .venv/bin/python tools/docs/verify_documentation.py --root . --strict -> "warnings": 0
$ bash scripts/verify_links.sh --root .                             -> PASS (1355 links resolved)
$ .venv/bin/python scripts/verify_docs.py --root .                  -> PASS (0 findings)
$ .venv/bin/ruff check scripts/verify_claims.py tests/test_retracted_figure_gate.py -> All checks passed

GENERATED-ARTIFACT NOTE
  `.aegis_ai_context/MANIFEST.json` hashes three of the swept documents
  (`DEPLOYMENT_GUIDE.md`, `docs/FAQ_TECHNICAL.md`, `docs/PLATFORM_OPERATOR_GUIDE.md`),
  so `tests/test_ai_context.py::test_manifest_is_deterministic_explicit_and_non_circular`
  failed on the first post-sweep suite run. Regenerated with
  `python scripts/generate_ai_context_manifest.py`; the diff is exactly those
  three sha256/size pairs; the test passes again (11 passed).

FOLLOW-UP AT CLOSURE (the sweep had to satisfy the strict doc gate)
  With the sweep in place, `tools/docs/verify_documentation.py --root . --strict`
  exited 1 on the corrected row itself:
    docs/FAQ_TECHNICAL.md:118 — "unqualified v4 external publication or release
    claim is prohibited in strict mode"
  Cause: that rule's trigger is `\bv?4(?:\.0(?:\.0)?)?\b` within 100 characters
  of `published|released`, and the bare `4` of "32.4 s" matched next to the
  phrase "(as previously published)" in the same row. The prohibition is right;
  the trigger is too wide, so the row was rephrased to "(as earlier reported)"
  AND the false-positive class was registered instead of patched inside this
  REG: `AUD-30` / `REG-D34` (the same discipline as `AUD-29` / `REG-D33`, opened
  because this REG's work surfaced a gap outside its scope).

FINAL GATES AT CLOSURE
  $ .venv/bin/python tools/docs/verify_documentation.py --root . --strict  -> exit 0, PASS, warnings 0
  $ .venv/bin/python scripts/verify_claims.py --root .                     -> PASS (102 claims, 0 findings)
  $ bash scripts/verify_links.sh --root .                                  -> PASS (1357 links resolved)
  $ .venv/bin/python scripts/verify_docs.py --root .                       -> PASS (0 findings)
  $ .venv/bin/python -m pytest tests/ -n auto -q                           -> 6896 passed, 118 skipped in 87.76s (0:01:27), exit 0 (shell, PYTHONPATH unset)
  $ .venv/bin/python -m pytest tests/test_ai_context.py -q                 -> 11 passed (manifest regenerated again after the rephrase)
