# Local validation — 2026-10-28

**The fictional Audit forward application failed substantive checks. Arithmetic and structural conformance checks pass locally. The full controlled behavioral benchmark is not yet run.** No deployment, new users, market-fit zone or outside usefulness is claimed.

## What changed

Audit now works from a dated source census into embedded methods, distinct promises and useful mechanisms, exact source-to-target ownership, install boundaries, research provenance, separate usefulness evidence and a reversible implementation sequence. Experiments and learning assets retain their own value. Confirmed owner decisions and applied visibility changes are separate from proposals. The rubric owns the scales, dispositions and report records; the longer conditional procedure is a portable reference.

## Checks run

- `python3 -m unittest discover -s tests -v`: 19 tests passed. This includes both preserved model-authored arithmetic failures, a mixed-cap regression, composition above a cap, duplicate check IDs, invalid values, lower-rung reasoning, all 256 need vectors against all 64 gate subsets, and CLI operation from a copied standalone folder. Eight structural Audit regressions check independent census/method coverage, existing paths, interfaces, revisions, fork buckets and numeric public placements. Their positive fixture is authored scaffolding, not a model or semantic pass.
- `python3 scripts/build.py`: built the static site and deterministic portable ZIP.
- `python3 scripts/check.py`: local links/assets/fragments and shipped Markdown destinations passed; no missing targets or duplicate IDs. The canonical skill/reference/script files and ZIP matched byte for byte. Three fictional placements were recomputed from the same rubric parser. Personal-information/credential hygiene scan passed.
- `node --check site/site.js`: passed.
- Browser inspection at desktop width 1238 and phone width 390: the Audit section rendered with readable headings, links and the disposition table. Document scroll width matched viewport width in both. The phone table retains its local horizontal scroll. Temporary viewport override was reset to desktop width afterward.

## Evidence limits

The [arithmetic regression record](evaluations/2026-10-28/arithmetic-regression/REPORT.md) keeps the earlier failed numeric outputs. Rejecting false arithmetic does not establish that a model's gate judgments are true, or that the Audit improves decisions.

The [single completed forward application](../evals/audit-forward/application-openai-2026-10-28/ASSESSMENT.md) used a frozen method packet and fictional sources. It covered all six repositories and respected the keep-public boundary, but omitted numeric placements and interfaces, proposed unsupported experiment/UI moves, and overstated install/license evidence. It failed. The earlier provider transport failure remains a separate record. No reroll replaced the failed response; this is not a controlled comparison or causal efficacy claim.

The later structural validator catches omissions, nonexistent paths and arithmetic errors, but cannot decide whether a merger, audience, research claim or license inference is sound. Engine None now distinguishes deterministic tools from AI expertise. Authorized private/local foundations remain independent scope/owners, not failed public products. No further model application of these changes is claimed.

The recipe was not edited. It was not executed or revalidated in a host during this change; no `recipes` tool was available in this environment. The optional mode body remains short and unchanged. No global install, new remote, push, publication or source retirement was performed.