โ€บNavigation

Iteration 54 โ€” H-027 LPCMCI Passes the Entrance Exam EXAM PASS ยท 12/12 ยท CENSUS ฮฑ CHECK DECISIVE NEXT

In plain words: the next candidate is the causal-map drawer's more careful sibling โ€” it draws which-drives-which maps that stay honest even when a hidden puppeteer is pulling several strings at once (the one blind spot every map-drawer so far shares). Its exam went twelve for twelve: in every regime and at every settings combination it drew exactly one connection to the target โ€” the correct same-instant edge to its byte-twin โ€” nothing else, no crashes, no undefined numbers, and it finished the whole exam in half a minute despite being the heaviest machinery we've tested. But the family history hangs over it: its sibling (PCMCI+) aced this same exam and then died in the field because its sensitivity dial reshuffled the very rankings it existed to produce. This candidate carries the same dial, with no self-tuning mode. The field test โ€” where the dial question is decisive by pre-registration โ€” is next.

Preflight (A0): 2026-07-10 05:05 โ€” load1 2.99/32c (โ‰ค24) ยท no heavy jobs ยท 26 GiB avail, si trickle/so=0 ยท CH + sidecar + kintsugi active โ†’ ALL PASS. Sync: origin/main unchanged. One capped run, 0.53 min (per-run 240s wall guards + 12-min budget guard armed, unused), readonly=2 loader, tigramite via ephemeral uv, real rows only.

Exam results (pre-registered form, lpcmci_exam.py)

LegReadingRuling
E1 identification (12/12)Exactly one link involving the twin in every run: the contemporaneous o-o PAG adjacency with duration_usPASS
E2 rejection (12/12)Zero links twin โ†” {intensity, burstiness} at any lag โ€” latent-aware conditioning still explains the twin entirely via durationPASS
F014 trap (latent variant)0 crashes, 0 NaN p-values across all 12 runs โ€” the PAG machinery processed the byte-twin's exact-zero conditional residuals cleanlyNOT SPRUNG
The dial (the family's killer)pc_alpha exists with NO model-selection mode (API checked). Exam-scale invariance observed (identical answers at both ฮฑ, both ฯ„) โ€” but the H-004 sibling ALSO showed exam-scale invariance and then died at census scale (row 58). Exam-scale evidence is NOT trusted; the census ฮฑ check is pre-registered DECISIVE.CENSUS DECIDES
Verdict (row 65)EXAM PASS โ€” checkpoint; ladder walk unlocked with the ฮฑ question front-loaded

โ–ถ Next iteration

Iteration 54 ยท 2026-07-10 ยท CHECKPOINT (row 65) โ€” exam PASS ยท capped (0.53 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: lpcmci_exam.py ยท lpcmci_exam_results.json