# Author validation

Date: 2026-11-04. These are author checks, separate from total-owner acceptance, deployment, publication and adoption.

## Frozen fitting and evaluation

Generation, protocol, configuration, audit and cases were frozen before model implementation and fitting. An independent scalar objective checks all 26 parameters of a tiny RNN by finite differences: maximum error 1.2051422360048747e-10. The recurrent Wh gradient is nonzero. Thirty-one explicit Adam history steps, immutable/zero clipping cases and a 65-bit target check pass.

One actual 80-epoch fit performed 2,720 updates, 307,760 representation visits and 2,052,000 bit visits. Gradient clipping was configured but triggered zero times. Validation selects checkpoint 80: all 554 integers and 3,679 bits correct, bit NLL .000803563992973165. PRETEST binds initial/selected arrays, all traces, gradient history, validation arrays, model, data and code before testing. A separate evaluator freeze precedes recurrent/reset testing. No trained code, weights or selection changed after test.

## Independent numerical checks

`verify_results.py` imports neither core nor generator. It reconstructs all 5,460 small representations, 2,080 unordered groups, roles, exact targets/carry, and exhaustive/random/padding/stress membership. Every 1,266,706 evaluation mode-bit and 3,679 selected validation bit positions is compared to scalar recurrence. Only identical complete histories share cached states: 185,563 history prefixes and four reset patterns. All 6,486,633 numeric comparisons pass; maximum error 8.526512829121202e-13 includes summed NLL. All 1,304 metric groups, 58,566 integer failures and checkpoint choices agree. Empty support remains null.

Node checks every independent prefix state and all 64,856 complete BigInt requests: 129,712 mode-case outputs and 1,266,706 bit positions. There are 2,542,083 numeric comparisons, maximum error 1.3322676295501878e-14, and 32 invalid-input/model/schema/nonfinite/sparse/forged/inherited rejections. Model immutability, a single-read getter input, asynchronous grouping snapshot, padding/minimum-width metadata and zero-tie behavior pass. A synthetic zero-output model remains wrong (`0` versus reference `16`), without answer repair. It is not a trained result.

Four controlled Node DOM/Web Crypto scheduling scenarios pass: Clear during the initial request, a newer request finishing first, invalid JSON suppressing an older valid result, and Clear during a large-input request. Changing the hidden view cannot revive a rejected report. This is a unit scheduler, not browser/network interception or native evidence.

The Python oracle generates the roughly 21 MB prefix cache. It is derived and reproducible, so the release omits it. Run Python verification before Node verification on a copy.

## Actual browser and reports

Official CUA used a hidden own browser and local server 17897. Thirteen final native downloads cover ten distinct ordered operand/width cases: six presets, an odd result above 2^53, explicit field recovery, post-Clear loading, operand reversal, minimum fitting width, and width-1 zero. A final form-isolation download repeats the field-recovery fixture byte for byte. The inspection select has no form owner; Enter/Escape after invalid JSON leaves the report cleared. Six preset downloads equal their actual DOM JSON bytes.

Every native step is independently replayed: 708 mode-bit steps and 5,664 hidden coordinates, plus logits, log probabilities, probabilities, NLL, bit truth, carry, hashes, input mode, grouping, padding and integer strings. Maximum scalar error is 3.552713678800501e-15. Maximum operands preserve `36893488147419103230`; `9007199254740993 + 2` preserves the odd result `9007199254740995`. Wrong reset answers remain wrong. Two initial exports predating padding metadata are retained as `native-initial-default*`, outside the final report count.

Twenty actual bad JSON/type/format/range/BOM cases clear reports and all bit/state rows, disabling export. Five field cases—blank operand, hex, 21 digits, blank width and width 65—also reject. JSON numeric operands are rejected rather than silently losing precision. A visible view change after rejection cannot restore old inputs. Explicit field editing restores field mode. Clear empties all inputs, JSON, cards, bit/state rows and current report; a fresh example then works. Both actual hidden-state tables agree with full report activations to four displayed decimals.

Actual 375/390 layouts have no page overflow; controls are at least 46 px and keyboard focus is 3 px. ArrowRight scrolls the bit region 40 px at 375 and a cumulative 80 px after the 390 check. Tables scroll internally rather than splitting integer digits. Both scientific PNGs load. Desktop and full mobile JPEGs were inspected. The viewport is reset; console warnings/errors are empty. No external user window, login, cookie, permission or public upload was used.

Full native files remain in the parent Projects directory. The release includes compact summaries and two complete examples. Native replay scripts require their recorded inputs; they do not substitute for a reviewer's independent browser checks.

## Findings and boundaries

Recurrent mode passes all 64,856 unique cases. Reset gets 6,290 integers right (9.6984%) and makes 256,470 wrong bits. In the 64,499-case primary union, reset gets 6,114 integers right (9.4792%). Complete stress includes 17 known fit representations; all overlaps and dependencies remain tagged.

The figures retain all 80 epochs, nine suites, 500 padding views, 127 carry probes and the default states. Padding raises bit accuracy without changing wrong integers. Reset uses the same weights, changes hidden-state distribution, and is not a separately trained baseline. There is no carry-neuron identification, calibration claim or universal arithmetic theorem. Original artifacts are Apache-2.0; standard mathematics is not claimed as an invention.
