# AI Harness + audit validation

Validated 18 September 2026 against `REPO-SNAPSHOT.json`; web uses `origin/dev-sunbeam`.

## Passed

- Census matches **all 16,246 pinned Git-tree paths** exactly (13,538 baseline; 2,708 secondary mobile). Modes, object IDs, explicit exclusions and branch separation checked.
- TS/JS/Python AST input coverage: **8,615 eligible files**, zero parse errors. Python uses 3.12. Discovery/classification regression tests distinguish runtime docs templates, developer skills, product skill scripts, tests, vendor files and sensitive configurations.
- 2,771 text candidate files; 151 baseline product invocation syntax candidates with false-positive categories kept explicit. These counts do not assert semantic or runtime completeness.
- `build.py`: seven chapters, **16 feature walkthroughs**, **511 entries** from **508 unique pinned files**. `build_audit.py`: 13 findings, 48 anchor files, full classified ledger and visual audit.
- `validate.py`: source text/line ranges/digests match Git objects; static IDs/links, baseline SHAs and web branch checked. Every static browser source also compared with the catalog by `smoke.cjs`.
- HTML fidelity regression: leading newlines preserved with `<pre><code>`. `workflow-globals` contains two literal NUL bytes, which HTML cannot represent exactly: no-JS view explicitly substitutes ␀; JSON and JS restoration preserve exact original text. This exception is tested.
- `validate-audit.py`: complete tree accounting, expected AST inputs, metadata-only sensitive exclusions, classification fixtures, all baseline backend product `SKILL.md` assets embedded, findings and static audit links/anchors checked.
- Browser smoke: all chapters/walkthroughs, SVG label containment, keyboard map navigation, selected source fidelity/deep links, full-text search, feature/type filters, reset/empty states.
- Audit browser smoke: repository totals, baseline/secondary/all scopes, 50-row pagination, exclusions, finding anchors, file details, SHA links, embedded-source links, sensitive metadata, gitlink disclosure, reset and empty states.
- Clipboard success against a stub and graceful denial; explicit offline export errors. HTTP(S) blocked throughout browser tests; no live services/models called.
- 390px containment for atlas chapters, selected source, audit ledger and file detail. No-JS chapters/static catalog plus audit findings/downloads remain available.
- Refreshed `harness-overview.pdf`: **one A3 landscape page**, via offline browser print. Refreshed four browser previews; desktop overview, prompt explorer and audit inspected visually.
- Original blueprint validator still passes: 14 views, 144 components, 131 unique pinned source paths. Original blueprint snapshot/PDF/ZIP preserved.
- Service working trees unchanged: five clean repositories; only pre-existing untracked `frontend/electron/package-lock.json`. Web working branch remains `prototype/cc-concept-jz`; its evidence comes from pinned `origin/dev-sunbeam`, not that checkout.
- Limited common-token/private-key pattern scan of all 511 catalog texts: no matches. This is not a comprehensive secret audit.

## Limits

- Exhaustive file accounting and recorded rule passes are **not** exhaustive semantic review. Finding anchors are source-reviewed; other files are curated assets or automated classifications. Generic transports, aliasing/reflection, indirect construction and non-TS/Python syntax remain incomplete analysis boundaries.
- No live configuration, tenant history/skills, external SDK preset bodies, deployment certification, model/prompt-quality evaluation, security penetration testing or retention/log inspection.
- The web `qa-aic` gitlink is recorded but not recursively fetched/audited. Vendor/generated/binary/sensitive files have explicit exclusion records, not prompt reviews.
- Successful online CDN-powered Copy-image / PNG / PDF exports are not certified. Source-copy uses a clipboard stub, not OS integration. Exporting a scrollable source pane captures its visible portion; Copy source / JSON provides full text.
- The source catalog and HTML embed internal code; share only with authorized readers. Secrets-pattern screening does not make these public-safe artifacts.

Browser tests require macOS Chromium execution approval and use existing `frontend/web/node_modules/@playwright/test`; no dependencies were installed. Reproduction commands are in `HANDOFF.md` and `AUDIT.md`.