Record the only defensible result: 2 of 9 gates pass, so NKE remains a research model with no deployment permission.
These readings update independently. They are not model inputs, deployment evidence, or a trading recommendation.
TECHNICALS THROUGH · TIMESTAMPED FALLBACK
The point-in-time fundamental ridge ranks first on development data, but the terminal holdout passes only G3 + G6. The honest recommendation is NO-GO: publish the result, keep the directional model switched off, and continue collecting unseen evidence.
Source hashes, next-session filing availability, actual ex-dates, terminal chronological isolation, explicit flatness and stability gates, exact shifts, and one canonical result file make the corrected run auditable and reproducible.
Improvement uncertainty crosses zero, AUC uncertainty crosses no skill, ECE is 10.7%, 6 of 10 folds improve, and 4 of five runs improve but dispersion is 1.10%. Forecast standard deviation clears 2%, while its P95–P05 span still misses 8%.
Candidate selection, tuning, scaling, and calibration exclude these terminal labels. The corrected protocol was frozen before this rerun, but an earlier provisional run inspected the same period; this is not independent confirmation or external preregistration.
95% block interval -0.003414 to +0.005436.
Bootstrap lower bound 0.448.
Holdout loss comparison versus the frozen prior.
Five-bin ECE; slope 2.868 and intercept -0.668.
Probability standard deviation; P95–P05 span 6.4%.
NO-GO · research only.
Freeze this exact pipeline and record paper-forward predictions before every NKE close. Reassess only after enough new outcomes accumulate to evaluate all nine gates. Any specification change starts a separately labeled model and receives its own future holdout.
The latest research probability is 48.7%. It is not a live signal, a forecast to trade, or evidence that NKE is more likely than not to rise in an economically useful way.
Continue through the audit-corrected Model 3 workflow.