# Local validation

Date:2026-11-04. This is author-side evidence for the frozen candidate, not independent acceptance, publication, free-cloud availability or adoption.

## Computational checks

`test_audit.mjs`:120 assertions, hand8-case fixture, declared hard/probability denominators, null support, abstentions with vectors, zero gold probability/infinite NLL, reported-label/argmax mismatch, typed missing-class collision, all duplicate occurrences, extras/blanks/malformed input, invalid probabilities, ASCII ID boundaries/Unicode, schema/UTF8/BOM/byte/physical-line caps, CRLF/terminalLF semantics, mutable caller-array snapshot,64MiB report cap and full3080 summary.

`verify_independent.py`:standard-library oracle independently reads original NPZ NPY float64 bytes, recomputes all237160 probability scalars through scalar stable softmax and matches every originalCSV argmax/confidence/p_gold/NLL/Brier row. Maximum vector discrepancy6.661338147750939e-16. Counter-derived original3080 hard scores/soft bins reproduce the report.225 exhaustive two-record combinations cover known/unknown gold and x/y/abstain/missing/invalid predictions. Complete input source text/bytes/SHA checked.

`test_ui_races.mjs`:four controlled promise interleavings in a Node DOM stub, explicitly not browser testing: Clear during fetch, newer preview winning older fetch, invalid newer file selection suppressing stale valid load while preserving active report, and Clear removing prior audit/preview DOM. No in-browser application-function execution or network interception was used.

## Actual official browser checks

A separate hidden own IAB and local loopback server used actual native file chooser for both fixture inputs, both3080-case inputs, empty prediction file and five global rejection cases. BOM, invalidUTF8,5001lines,>10MiB and malformed inventory rejected without replacing active metrics. Final Clear removes source/card/case/line/report/preview DOM; after Clear, a new demo can apply and download normally.

Five final source-attributed actual native downloads (earlier pre-attribution exports superseded): fault, all3080, valid empty predictions, fault with display filter invalid, and post-Clear fault. Full fault native bytes equal displayed JSON and core report (15959bytes). All3080 native report equals full core report (30597628bytes), SHA2563e624b54741351b6ac9d7e44474b59ad1ba58702635ceda03435c95265e0ad2f. Filtered and post-Clear downloads still equal the entire8-case/11-line fault report. Empty file report retains3080 missing cases, zero coverage, null conditional/probability scores, completion bounds[0,1]. Raw files live outside the release; summarized hashes/counts in UI-EVIDENCE.json.

Actual display status/search and inventory pagination (124pages), true-class selector including last category, expandable per-class/bin/physical-line tables and line pagination were exercised. Filters do not change exports. Script-like sample text renders literally, no image nodes. Keyboard Tab reaches labeled file input with3px visible focus. Actual375/390 viewports have equal document/viewport widths after fixing long count text wrapping; inputs/selects≥46px. Screenshot previews visually inspected. Viewport reset. Console warnings/errors empty. No visible external-screen window, permissions, login, paid service or public upload.

## Provenance and limits

Original contract freeze precedes core implementation. Before core, a single synthetic hand ECE arithmetic typo .15 was corrected to .35 and recorded separately; formula/bin conventions did not change. Original freeze/erratum both retained. Earlier mobile check caught410px content overflow; CSS now wraps count paragraphs. Clear initially left hidden old DOM nodes; final code removes them, with controlled and actual browser checks. These are author-side fixes before candidate freeze, not reported independent PASS.

Known TinyHash/BANKING77 outputs are recalculated, not new inference or training. No blind test, model improvement, calibrated per-case confidence, independent gold quality, dataset independence or adoption demonstrated. Zero handling/column order/binary scaling/bin edges are explicit conventions, not numerical identity with every library default. No external account/browser/private content in candidate. Future candidate edits require scoped re-review.

## Candidate boundary amendment

The first47-file staging inventory was retained read-only before handoff. Final staging-v2 adds standalone method references and per-source exact bundled-byte attribution (otherwise null/unspecified origin). Metrics/input conventions unchanged. All120 core assertions,225 independent state fixtures,237160 original-logit scalars and all five native reports were checked for the final core. User source hashes establish matching bytes, not authenticated authorship or license assurance. Only staging-v2 is submitted for independent review.
