Review decisions, with evidence attached
What was accepted, what was already corrected and what remains unverified in the supplied scientific review.
What the review changes
The added review mostly assesses 0.4.1. The implementation baseline was 0.6.0, which had already corrected the SI mapping, source slugs and several reporting issues. This release addresses the remaining predictive-test and status-reporting gaps.
| Finding | 0.7.0 response |
|---|---|
| E01-S contrast is too narrow as the flagship | Prediction on new independent specimens becomes primary; null contrast is secondary. |
| Conventional softening must be taken seriously | Runnable synthetic reference and explicit material/rate limitations; calibration still required. |
| Primary metric and uncertainty need a concrete implementation | Scorer, simultaneous intervals, calibration-draw interface and error/leakage checks supplied. |
| Measured AICc weights should be visible | Displayed directly from the unchanged four-model archived results. |
| External completion claims lack supplied records | Unverified claims remain unverified; exact evidence required is listed. |
| Same-author code agreement is overstated | Labeled an implementation cross-check, not independent scientific validation. |
Preserve the result that is actually reproducible
573 species; 372 training, 201 held out; 99 families. The 10 kg hinge primary RMSE is 0.200464 dex versus 0.198095 for a single line. Better training fit does not establish better held-out prediction. The unsupported 626-species and “curvature confirmed” summary is not substituted for this result.
Implemented is not experimentally confirmed
Twelve new implementation checks and three synthetic decision controls run successfully. The conventional reference parameters and all scorer fixtures are synthetic. Physical candidate/rival calibration, justified experimental N, prospective registration and independent reports remain open.
Full response and evidence requirements · Exact checks · External status