# Author validation scope

Protocol/config/source/research were frozen before implementation and fitting.
Six actual169-parameter fits ran120epochs each:720epochs,19,440updates and606,960
row visits. Three saved initial weight sets are shared across paired seeds, with
separate pool-specific normalization and training. PRETEST preceded the frozen
evaluator and first current-model test. Original source values were already
audited; no historicalblindness or third-party timestamp proof is claimed.

## Math and complete denominators

Before fit: independent11-parameter scalar MSE+L2 finite-difference gradients,
max3.993e-11; zero-output hand loss;17stepAdam scalar recurrence; Decimalzero and
exponent normalization; exact pools/commoncases/train-onlymeans checked.

After test: separate scalar oracle imports no core math. NumPy only reads NPZ.
It reconstructs all1,503source records, label-free grouping/pools, train-only
means/populationstd/targetscales, membership/catalog metadata, all2,850selected
neural forward records, all3,800method rows,392overall/group metrics,603paired
test rows and all237B-worse rows. Collection membership/uniqueness is checked,
not only row counts. Largest scalar prediction difference2.843e-14dB.

PortableJS checks all475validation/test contexts × six models, all input/condition
hashes/catalog exposure/normalizers/means/predictions, max2.843e-14dB;20badinputs
reject.11malformed models,8malformed catalogs, forged objects, inherited modelID,
duplicate/incomplete catalog source rows and stale mutable-input hash snapshots
reject. Frozen model/catalog arrays cannot be accidentally edited. A synthetic
valid large-bias unit fixture returns over1,000dB, demonstrating no hidden cap to
the original target range; it is not trained weights or a benchmark result.

Numeric hash keys expand finite scientific notation to plain decimal, collapse
negative zero and preserve JavaScript's parsed numeric value. This matches the
original Decimal source key convention without claiming original JSON-byte
identity. Unit fixtures include1e-7 and the smallest positive numeric value;
an actual browser input with angle1e-7 exports0.0000001in the condition key and
passes an independent Python Decimal/scalar replay. No input truncation or angle
binarization occurs.

## Actual hidden browser

Official trusted IAB loaded the exact model/catalog byte SHA and hashed default
sample. Default is the first common-test original ordinal29, seed17, chosen by
protocol rather than score. Its B prediction is worse than A relative to the
source target; this remains visible. Source target is metadata only.

Final20valid native submissions cover fixedtest/base/validation tuples, joint
combinations at all marginal minima/maxima, fractionalfrequency/angle, numeric
field editing after invalidJSON, and tinyrealangle. Allthree seeds are actually
selected for the six main cases. Every report includes allsix outputs, not just
the two displayed cards; all120point outputs, normalizers, baselines, condition
lookup and numerical hashes independently scalar-replayed. Max2.843e-14dB.

Fifteen native invalidJSON cases cover missing/extra fields, bool/string/null,
array, BOM, bytecap, badJSON, numeric overflow and eachfeature's range guard.
All reject with old result, report text and download href removed. Changingseed
does not silently switch invalidJSON back to old numeric fields. Separately,
blankfrequency and blankangle reject rather than becoming0; an explicit numeric
field edit switches mode and validates the current fields. Clearinputs empties
allfive fields and export, and Loadfixed restores the source case. Seed changes
clear old results. Revision guards bind async reports to their input snapshot;
testing is bounded, not an exhaustive concurrency proof.

Two final reports were actually downloaded through native Blob links: recorded
source and custominput. Full report bytes match DOM text. Both preserve feature
order/explicitunitmap, allinputs, six outputs/epochs/pool exposure/normalizers,
model/catalog/source SHA, originalauthors/DOI/CC BY4/codeApache/NASAref and user
inputterms. The20case collection is DOM snapshot evidence, not20native downloads.
Reports authenticate neither inputs nor physical experiments.

Desktop and375/390mobile screenshots plus both full standalone figures visually
inspected. Page client/scroll widths equal requested widths; numeric inputheight
48.39px at390. The metric table scrolls within its own container. Allfigureassets
load, consolewarn/errorlistempty, viewportreset. No externaldisplay/nativewindow.
This is bounded layout/control checking, not fullaccessibility or privacy proof.

## Interpretation and release

Allsix testMAE/RMSE and all237unfavorable paired rows remain. Three paired-seed
results reuse201testcases; not603independent cases, confidence/significance or
ensemble estimates. Swapping185rows changes trainingcoverage/statistics as well
as conditionexposure; no isolated causal effect. Exact tuples are proxies, not
runIDs. Bmetrics do not estimate unseen-condition generalization. OutputscaledSPL
dB is not dBA/pressure/energy or engineering/noisecompliance assurance.

Original UCIvariables/source values are unchanged; angleBinarycarddiscrepancy is
described, not silently coerced or claimed fixedupstream. Code/docsApache2 and
source/weights/results/figuresCCBY4 scopes remain separate. NASA record/abstract
read only; no reportalgorithm/code/figure reproduction. User inputs keep original
terms. No employer/private/competition material, remoteupload or newpermissions.
Independent total-output sixfreshfits and native product review remain required.
Later dailyHF additionally verifies actualfreeStatic/build/App/model/input/report;
local passes are not publication or adoption.
