Frozen evaluation
47 chronological origins · 26 exact global reset events · 7 positive 24h outcomes · 13 positive 48h outcomes.
MODEL RESEARCH
CNR-0.4.2 research results
CNR-0.4.2 is the accepted next-generation research architecture for predicting a broad/global Codex automatic reset. It is still research-only and does not replace the public CNR-0.2 forecast.
24-hour result
CNR-0.4.2 Brier: 0.133382 vs baseline 0.133834. Log loss: 0.462051 vs 0.465475. Brier skill: +0.34%.
48-hour result
CNR-0.4.2 Brier: 0.210662 vs baseline 0.221653. Log loss: 0.625665 vs 0.666093. Brier skill: +4.96%.
Why 0.4.2 was accepted
The locked candidate beat the same-target baseline on Brier and log loss at both horizons, both model heads converged on every origin, temporal features stayed disabled, and the retained 48h regime family passed its ablation gate.
ARCHITECTURE
24h stays simple; 48h keeps the useful regime signal
The 24-hour head uses shared historical/context features with no recent-regime features. The 48-hour extension retains gap compression, recent-rate lift and short-gap persistence. The change was predeclared before the locked rerun.
BASELINE
Compared against the same target
The benchmark is an empirical intercept-only 12-hour hazard trained on the same broad/global target and chronological training rows. This avoids claiming skill by comparing against a model that predicts a different event definition.
LIMITS
This is not a production-accuracy claim
The sample is still small, coverage remains incomplete, and the formal calibration gate has not been met. Product activity, targeted-reset context, primary hints, secondary rumors and negative evidence also lack enough historical variation to claim that they improve forecasts.
FORWARD SHADOW
Future forecasts are scored automatically
Only canonical 00:00 and 12:00 UTC pre-announcement forecasts count toward forward scoring. Mature 24h and 48h horizons are resolved against the reviewed global-reset ledger and compared with the frozen same-target baseline using Brier score and log loss.
Current CNR-0.4.2 shadow forecast
— within 24h · — within 48h
Loading research shadow…
Forward evidence gate
Loading shadow duration and resolved cycles…
24h forward score
No resolved score loaded yet.
48h forward score
No resolved score loaded yet.
Research status: CNR-0.4.2 is research-only, publicationApproved=false, and production-write-forbidden. Promotion requires additional forward evidence and a separate human-reviewed change.