โ€บNavigation

Iteration 55 โ€” H-027 LPCMCI EXCLUDED KILL โ€” ฮฑ-FRAGILE 0.393 (SIBLING: 0.625) FAMILY LESSON ร—2

In plain words: the careful sibling met exactly the fate we pre-registered as decisive โ€” and failed harder than its brother. Its job would be to rank features by how much their causal wiring changes between regimes; turn its sensitivity dial between two perfectly reasonable settings on identical data, and that ranking survives only 39% intact (its brother managed 62.5%; our bar is 90%). The extra machinery that makes it honest about hidden puppeteers โ€” deciding which way each connection points under uncertainty โ€” turns out to AMPLIFY the dial's influence rather than tame it. That makes the lesson a family law, measured twice: on this market, wiring-map rankings from this school of methods are threshold-brittle at scale, and acing the entrance exam predicts nothing about it (both siblings went twelve-for-twelve before dying at the same gate). The repair path is on record for both: versions where the dial tunes itself. Next in the queue: a completely different school โ€” a map-drawer built as a deterministic optimizer with NO sensitivity dial at all, which is precisely the property this family lacked.

Preflight (A0): 2026-07-10 05:25 โ€” load1 2.72/32c falling (โ‰ค24) ยท no heavy jobs ยท 28 GiB avail, si trickle/so=0 ยท CH + sidecar + kintsugi active โ†’ ALL PASS. Sync: origin/main unchanged. One capped run, 4.9 min, readonly=2 loader, watchdog clean, resume grid saved, real rows only.

Census results (pre-registered rule in lpcmci_m3_census.py โ€” row-58 design, honest sibling comparison)

CheckReadingRuling
The DECISIVE ฮฑ checkstruct_instab (the role readout) dial correlation = 0.393 โœ— (bar 0.90; sibling PCMCI+ read 0.625, row 58) โ€” the PAG orientation phase AMPLIFIES threshold brittleness. link_rate: 0.957 โœ“ (not the role readout)RULE 1 FIRES โ€” the sibling's fate
Degeneracy / redundancy (recorded)Not degenerate (struct_instab q10/50/90 = 0 / 0.43 / 0.72); not redundant (worst |ฯ| = 0.695 vs HSIC-X min_p)kill is the dial alone
Verdict (row 66)EXCLUDE(ฮฑ-fragile at census scale โ€” 0.393 < 0.90 on identical rows; row-58 precedent applied as pre-registered) โ†’ BONEYARD, re-proposal only with the dial derived away
THE FAMILY LAW (constraint-based causal discovery on this substrate), MEASURED TWICE:

              entrance exam        census ฮฑ check (role readout)
  PCMCI+      12/12 PASS   โ”€โ”€โ”€โ–บ    0.625  โœ—  (row 58)
  LPCMCI      12/12 PASS   โ”€โ”€โ”€โ–บ    0.393  โœ—  (row 66 โ€” orientation phase amplifies)

  โ‡’ exam-scale dial invariance certifies NOTHING about census-scale dial stability
  โ‡’ wiring-map rankings from this school are threshold-brittle at scale here
  โ‡’ repair path (both): the dial must tune itself (pc_alpha=None / ฮฑ-integrated readouts)

โ–ถ Next iteration

Iteration 55 ยท 2026-07-10 ยท EXCLUDE (row 66, BONEYARD +1) โ€” kills are wins ยท capped (4.9 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: lpcmci_m3_census.py ยท lpcmci_m3_census_results.json ยท lpcmci_census_grid.json