โ€บNavigation

Iteration 68 โ€” H-073 MMD M3 CLEAR ALL FIVE CHECKS PASSED ยท M4/M5โ€ฒ FINALE NEXT

In plain words: the kernel-distance candidate faced the field test that has killed six candidates this campaign: measured across all 130 features, does its reading merely repeat something a certified instrument already says? It does not โ€” and the most important miss is the one this batch was built around. Against the incumbent regime counter (the single instrument the FRAGILE-vs-CONDITIONAL boundary currently rests on), the correlation is a weak negative: the features whose value distributions shift most between regimes are NOT the ones the counter flags as verdict-unstable. Two instruments that disagree about which features are regime-sensitive is exactly what a second boundary instrument is for. Its closest relative in the stack (a moderate 0.84, still under the kill bar) is the evidence-magnitude readout of the e-value family โ€” regime heterogeneity tracks how much evidence accumulates against invariance, an interpretable kinship, not a repackaging. The reading was also recomputed two independent ways and matched to the last bit. One test remains: does it actually improve held-out forecasts?

Preflight (A0): 2026-07-10 13:05 โ€” load1 3.05/32c (โ‰ค24, falling vs load15 3.73) ยท no heavy nasimubd jobs ยท 25 GiB avail, si/soโ‰ˆ0 ยท clickhouse-server + opendeviationbar-sidecar + kintsugi + fxview-sidecar all active โ†’ ALL PASS. Sync: main pulled ff-only (+1 file); worktree 0 behind. One capped census run (1.38 min) + one capped verification (0.4 min), readonly=2 loader, deterministic, permutation-free.

M3 census โ€” five pre-registered checks (mmd_m3_census.py, 130/130 subjects)

CheckReadingBarRuling
1 ยท NSUB dial law (census scale โ€” the rows-51/53/58/66 lesson)ฯ(readout@1000, readout@1500) = 0.995≥ 0.90PASS โ€” subsample-stable
2 ยท k_v incumbent shadow (DECISIVE โ€” row-78 named adjacency)ฯ = โˆ’0.35 โ€” weak and negative: distribution-shift ranking disagrees with verdict-instability ranking|ฯ| < 0.95PASS โ€” not the incumbent's shadow
3 ยท Distributional cousinsIPP โˆ’0.679 ยท calgap20 0.786|ฯ| < 0.95PASS
4 ยท Corpse affinity (rows 70/72 fate-transfer law)gdro gap 0.641 ยท verex var 0.657|ฯ| < 0.95PASS
5 ยท Stack redundancy (full certified grid)worst |ฯ| = 0.842 vs eicp_log_min_e (evidence-magnitude kinship โ€” interpretable, not a shadow); all others ≤ 0.804|ฯ| < 0.95PASS
Verify-before-reportCensus cached-self-term algebra recomputed independently with the sealed exam's mmd2_unbiased on 2 features ร— 45 pairs โ€” bit-exact match both timesVERIFIED
Verdict (row 79)M3 CLEAR โ€” checkpoint. M4 held-out lift + M5โ€ฒb direction next, with the temporal-stability watch (row-70 law)

Dial audit: bandwidth derived per feature (median heuristic, row-45 convention) ยท NSUB inherited verbatim from the sealed exam AND swept per check 1 ยท estimator paper-native unbiased-U ยท grid = the frozen 10-env census grid. No free dials. Readout spread: 10th pct 0.0003 ยท median 0.030 ยท 90th pct 0.270 โ€” three decades of dynamic range, no degeneracy.

โ–ถ Next iteration

Iteration 68 ยท 2026-07-10 ยท CHECKPOINT (row 79) โ€” H-073 M3 CLEAR ยท capped (1.38 + 0.4 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: mmd_m3_census.py ยท mmd_m3_census_results.json