โ€บNavigation

Iteration 33 โ€” H-019 IPP Passes the Entrance Exam EXAM PASS ยท DISCORDANCE PROBE NEXT

In plain words: the next candidate asks a subtler question than any instrument so far: not "does the average relationship hold across regimes?" but "does the entire shape of the prediction hold โ€” including the tails, where financial risk actually lives?" A relationship can keep its average while its spread doubles; every SSR-based instrument in the stack is blind to that. The exam form: score each prediction with CRPS โ€” a strictly proper scoring rule that rewards getting the whole distribution right โ€” and test whether scores stay level across the ten regimes. It passed cleanly: perfect answers accepted (the scoring rule handles the exact-zero-error case gracefully, where a log-score would blow up), wrong answers rejected with the strongest separation any exam form has shown, exactly the right variable identified. The real fight is next: since this candidate and the reigning identifier both start from the same regressions, the redundancy gate must find the features where the shape shifts while the average holds โ€” that population is this candidate's entire reason to exist.

Preflight (A0): 2026-07-11 โ€” load1 4.11/32c (โ‰ค24) ยท 16 GiB avail ยท CH + sidecar active ยท no heavy jobs โ†’ ALL PASS. One capped run (0.49 min), complete-case real rows, permutations of real values only. Exam form pre-registered in-script.

Exam design (pre-registered)

ElementChoice
Candidate mechanismIPP (Henzi et al. 2024 Biometrika): invariance of the WHOLE predictive distribution via proper scoring rules. Bounded form: Gaussian predictive (pooled OLS mean + residual sd) ยท closed-form Gaussian CRPS per observation ยท statistic = per-env heterogeneity of mean scores
Null skeletonv5 fold-product unchanged (J=5 env-stratified folds, K=99 row permutations of fixed per-row scores, degenerate guard, e<20 accepts) โ€” only the statistic differs from the admitted oracle
PASS criteriaE1 truth-containing subsets accepted ยท E2 wrong subsets rejected ยท identified = {duration_us}

Result

Subsete_totalRead
duration_us (and all 3 supersets)1.0 (5/5 degenerate folds โ€” exact zero error; CRPS โ†’ |yโˆ’ฮผ| = 0, no log-score infinities)ACCEPTED โœ“
intra_intensity_qps17,287REJECTED โœ“
intra_burstiness31,816REJECTED โœ“
intensity+burstiness26,632REJECTED โœ“
Identified{duration_us} โ€” exactly right. EXAM PASS (row 42), 0.49 min

The decisive question, recorded now: IPP and e-ICP-fp both ride pooled OLS residuals. The M3 walk must include a discordance probe โ€” the features where IPP rejects invariance while e-ICP accepts it (variance/tail shifts at a stable mean, invisible to SSR). That population is the candidate's entire role case; if it is empty, the CRPS layer is decoration and the candidate dies redundant.

โ–ถ Next iteration

Iteration 33 ยท 2026-07-11 ยท CHECKPOINT (row 42) โ€” exam PASS ยท capped ยท zero synthetic data ยท append-only ยท evidence: ipp_exam.py ยท ipp_exam_results.json