CIF v0.21.3
Frozen candidate models tested once on a separate asset holdout. No grid search, no post-result mutation.
candidate-holdout-alt-v1-2026-08-07
Independent asset replication
No fitted coefficients
Exact current MOS control.
post-discovery risk: noneEqual-weight core; drops Relative Strength and Positioning without fitting coefficients.
post-discovery risk: moderateMedian of the same four core components; robust to one extreme component.
post-discovery risk: moderateTests whether discovery-negative Relative Strength and Positioning have reversed semantics.
post-discovery risk: highCross-sectional + per-asset replication
Date-matched spread
Median 30d excess-BTC spread between the two highest-score and two lowest-score assets on each date.
2.25 ppMedian 30d excess-BTC spread between the two highest-score and two lowest-score assets on each date.
1.15 ppMedian 30d excess-BTC spread between the two highest-score and two lowest-score assets on each date.
0.79 ppMedian 30d excess-BTC spread between the two highest-score and two lowest-score assets on each date.
0.11 ppEligible for forward confirmation?
What happens after this holdout
- Holdout basket was frozen before holdout results were observed.
- No grid search, coefficient optimization, or candidate mutation is performed after holdout results.
- Any nominated candidate becomes contaminated by this holdout and must be frozen before a new forward confirmation sample.
- This lab validates MOS hypotheses only; it does not validate AQS or Risk.
- No production gate changes in v0.11.
- The holdout results may nominate one fixed candidate, but may not edit it.
- If no candidate passes, we return to hypothesis design instead of lowering the gate.