# Author local validation

Date2026-11-04. Author-side evidence only; independent acceptance, deployment, publication and adoption remain separate.

## Frozen source, training and test

Original UCI ZIP/three members verified; all6497records/5318exact groups/1179duplicate excess rows/2cross-color groups/0conflicting-gold groups accounted for. Input-only group hash and label-blind source-ID selection reproduce923fit/335ownval each, full341red+982white test,2016unusedwhitefit/642unusedwhiteval. AllsourceDecimal numeric keys also match their float shortest-string roundtrips; no source-key alias repair required.

Protocol/config/source/audit/preparer frozen before model implementation/fits. Independent13scalar finite-difference gradient parameters max2.5294044636581248e-11,23explicitAdamhistory steps,313parameterarchitecture/train-only zero-std tests passed beforefit. Training code bound separately before first original-data gradient. Six actual100epoch fits,600epochs/9000updates/553800visits, three paired initial sets; full600ownsourceval traces and sixselected checkpoints. PRETEST before first current-model test binds all selected models, source, catalog, membership and code. Separate evaluator freeze before first complete test. No trained code/weights/selection changed after test.

Independent verify_results.py does not import core/prepare. Rebuilds full source/group/sampling/train-onlystats and recomputes9948selected-val/test neural forwards by scalar math, max8.881784197001252e-16. All10584method rows,192overall/qualitygroups including empty-supportnull,3969paired rows and both complete failure files, all600trace/checkpoint selections agree. This is regression; every continuous error is retained. Descriptive in-domain-worse1585/cross-domain-worse2384/equal0; no independence/CI claim.

Portable JS checks all1993unique selected-val/test input contexts and9948expected neural predictions, same max8.881784197001252e-16;39rejection/immutability checks include schemas/nonfinite/dimensions/sparse arrays/extra fields/missingmodels/hashmetadata/seed/inputbounds/inherited identities, forged unverified model/catalog, caller mutation and sourcecatalog snapshot during asynchronous hashes. Synthetic large-bias unit produces >10prediction, proving no hidden clipping; it is not a trained score. Tiny1e-7/negativezero/1e21/Number.MIN_VALUE canonicalization checks included.

## Actual product interaction

Official hidden own CUA tab and local server, no direct browserHTTP/CDP/Playwrightshell.26actual native complete JSON downloads: firstred/white test/fit/validation at each seed (18); all-marginalmin/max, fractionalalcohol10.125,tinycitric1e-7 (4); fields restored after badJSON/seed; postCleardefault; both cross-color exact groups (2). Every named source report equals displayed full JSON. Independent scalar replay of all156exported model outputs matches within8.881784197001252e-16; every source/model/catalog/config hash, allnormalizers, inputorder/mode/key/sha, selectedseed/epoch, exactgroup/fitflags/marginalflags/means/CCBYsource attribution matches.

Fifteen invalid JSON cases clear current report and disable export. InvalidJSON then seedchange cannot revive oldfields. Blankcitric rejects instead of becoming0; hexalcohol rejected as nondecimal. Explicit numericfield edit restores fieldmode and removes loadedsourceclaim. Clear actually blanks all11inputs/JSON/currentreport/sourcecontext and disablesdownload; a newrecord canload afterward. Invalid sourceIDred:9999rejects. Both cross-colorgroups return allmatching records, retain a shared initialrole, and never claim measurement authentication.

375/390actual documentwidth equals viewport, every input/select>=46px. Desktop/fullmobile previews visually inspected. Early mobile table layout split seeds/decimals; final tables have readable minimum widths inside keyboard-focusable horizontal-scroll regions with a visible hint. Actual ArrowRight moves the six-output region40px at375, no pageoverflow;3pxfocusvisible. Two full scientific figures load and have full-size links, display all600traces/3969paired differences/7938neural outputs and analytic baselines. Figures visually inspected. Final captures are JPEG in correctly named .jpg files; scientific figures are PNG. Viewportreset; consolewarnings/errors empty.

Native raw26reports and fullbrowser records remain in parent Projects for reviewer; compact hashes/check summaries and two complete real native examples are in release. verify_native.py requires the26parent files when run; it is not a substitute for independent actual browser tests.

## Limitations and negative findings

The white-trained NN on red holdout is worse than its analytic fit-mean baseline (.746975vs.636872MAE). Both main table and full scientific figure display the baselines. Fixed defaultred:0002also favorscross-domain atseed17 and remains. All1585in-domain failures preserved. Both domains' mean MAEs favor in-domain models; equalrows/updates/pairedinitials do not isolate causal shift. Source labels/support/normalizers/sourcevalidation/duplicateweights differ. Quality is ordinal sensoryscore treated numerically for loss, not a physical interval quantity. Entire source was already audited beforeprotocol, no historicalblindness or futurepopulationclaim. No calibrated perinputuncertainty, tasting/health/certification/engineering/adoption claim. Data/weights/resultsCC BY4, owncodeApache2.
