# Methodology and review boundary

Historical suite: ACCEPTED_OFFLINE_ONLY. The predecessor page, historical reconstruction and predecessor browser QA passed bounded independent review. The Jev prerequisite predecessor also received independent local scientific and desktop/mobile browser acceptance. This practical successor awaits independent scientific/editorial review and fresh browser QA; no acceptance transfers automatically.

The generator checks pinned input hashes and review receipts, reads the accepted aggregate and prose artifacts without rebuilding predecessors, and preserves every historical table cell. Supplemental endpoints retain accepted-source provenance rather than being mislabeled raw replay verification. The historical inventory remains 6,796 tasks (6,793 valid and three invalid); historical reserves include two separate smoke attempts.

New Jev data comes only from the independently audited aggregate export. That audit reconstructed native raw receipts, frozen actor/oracle labels and historical response records; hosted comparator aggregates were cross-checked against accepted reports, not newly raw-revalidated. Invalids stay assigned; probability diagnostics exclude invalids without imputation. Development cases share rules/trajectories; no independent fresh-trial or general regression claim is warranted. Renderer/provider/time confounds prevent causal attribution.

Only reading and one_step are complete for Jev 1.14. All other families are unrun/unavailable; null is not zero. Usage is an observed subtotal excluding the separate compatibility probe. Allowance is not billing. Phase timers exclude pacing/startup and durable journaling; no controlled speed comparison follows.

Run `PYTHONDONTWRITEBYTECODE=1 python3 research-release-practical/verify.py` from the private study root. This writes only the successor directory. It runs no provider calls, upstream generators or shared ledger writes and reads no credentials. The public aggregate bundle lacks private raw inputs and is not a standalone raw-replay package. Serve only public/, never the internal evidence directory.

Relative downloads are stable local links. Public evidence/protocol URLs, distribution clearance and external publication authorization remain pending. No external publication occurred.

## Practical-guidance evidence boundary

This successor adds interpretation, plausible applications, training guidance and proposed FS acceptance criteria. It does not add empirical observations, live inference, fine-tuning results or a production decision. All historical and Jev numerical tables remain unchanged. Recommendations are excluded from measured JSON/CSV fields. The source announcement is linked publicly for licensing and the announced fine-tuning route only; general availability and study-specific training gains are not established.

The proposed advisory 20% time reduction, tagging 95% precision and automatic-rejection below-1% miss targets are new recommendations to be reviewed against deployment losses, not frozen experiment gates, achieved findings, universal standards or authorization. The historical 95% CA computation gate and historical news calibration policy are unchanged. Future evaluation would require independently labeled chronological/event-grouped samples, real-workflow comparisons, repeated windows, uncertainty and subgroup checks. Creator and validator remain separate. No publication or production action was taken.