CHANGELOG.md

## 0.6 — 15 October 2026

Conservative out-of-support promotion policy; true browser inference from the synthetic scikit-learn forest; 204 numerical parity probes; source browser and Git mirror; complete English submission. 52 Python and 24 Node checks passed. Visual/native browser QA unavailable. No paid customer or accuracy guarantee.

# Internal prototype changes

## Package 0.5 — 15 October 2026

- Add standalone reproducible synthetic audit and complete JSON evidence, preserving the unchanged model and browser source.
- Document the 17/25 ordinary-row flag limitation without treating synthetic counts as production accuracy.
- Prepare a DEV-template article draft and a factual two-minute demo plan; no recording, public link, user feedback or submission is claimed.
- Correct eligibility documentation using the organizer reply: local scikit-learn qualifies in principle, eligible Uzbek residents can receive prizes, payout via Ramp; remaining rules and individual bank route still need checks.

## Package 0.4 — 15 October 2026

- Add original local browser report-viewer source using the existing precomputed synthetic report; no browser inference, uploaded user data or Alexa integration.
- Include review/confirm/export/reset gates and full-report SHA-256 binding. Async review/reset races cannot resurrect stale confirmation.
- Add 18 Node logic/static tests, all passing. These do not validate browser rendering, DOM interaction or downloads. Browser QA skipped because the required control-browser capability is unavailable in the managed environment; no preview server or alternative browser was used.
- Model 0.2, CLI adapter 0.1, original input/output fixture, dependency requirements and license notices remain unchanged. No external publication, registration, submission, API, account or cost.

## Package 0.3 — 15 October 2026

- Add typed CLI conversation adapter 0.1 around real ReviewQueue model output, not a stub.
- Bind demo confirmation to the full report snapshot: original rows, flags, scores and metadata survive export. No record correction or deletion.
- Add review/confirm/export/reset flow, strict pre-confirmation refusal and exclusive-create export. Review/reset invalidates confirmation.
- Add 20 synthetic adapter/integration tests alongside 28 existing model tests. No change to model engine 0.2, dependencies, license notices or the six-record fixture.
- This is an internal CLI prototype, not an Alexa connection, browser demo, public release, contest registration or submission. The historical HF26 archive filename is retained; this is not a new DEV project with an invented start date.

## 0.2 — 15 October 2026

- Reject duplicate JSON object keys, including nested and escaped keys. Duplicate record rows remain retained and flagged.
- Reject JSON floats whose decimal value would be silently rounded, overflowed or underflowed by decoding. Exact decimal strings are the recommended input.
- Encode output as UTF-8 before creating a file, so an invalid surrogate cannot leave an empty result file.
- Add 10 regression tests; all 28 tests pass. The original 6-row synthetic fixture still produces 6 rows and 4 review rows.
- Source bytes remain unchanged; output formatting and numeric spelling are not a byte-identical copy of the source.

## Reproduction before the fix

The existing CLI accepted both a duplicate `records` key and an extra numeric field `1.0000000000000001`, losing a value or precision during parsing. A JSON string with an unpaired surrogate was rejected only after an empty output file had been created. Regression tests reproduced these failures before implementation; the same cases now reject input before any output file is created.

No client material, external submission, account creation, payment, new dependency or downloaded model was involved. This is still an internal synthetic prototype; no production accuracy or competition eligibility is established.