โ€บNavigation

Iteration 56 โ€” H-030 DYNOTEARS EXCLUDED at the Door EXAM KILL โ€” WRONG, NOT BLIND

In plain words: the dial-free hope died fastest of all โ€” at the entrance exam itself, and in the most instructive way. Two problems, both fatal. First: handed the one question with a CERTAIN answer (the byte-twin's connection strength is exactly 1.0 by construction), its reading ranged from 0.97 down to 0.56 depending on its penalty setting โ€” turns out "no sensitivity dial" was true, but its penalty knob biases the answer by up to 44%, underived. Second, and worse: it drew up to THIRTEEN phantom connections for the twin โ€” a variable that has exactly one true relationship in the universe โ€” scattering the duplicate's signal across imaginary structure, where the two map-drawers we killed earlier at least drew exactly one edge every single time. It never crashed and never produced an undefined number: it fails by being confidently wrong, not blind โ€” which is precisely what the entrance exam exists to catch, at the cheapest possible price. Structure-learning is now zero for three on this market. Next: a different question entirely โ€” the worst-regime-vs-average-regime gap meters, examined as a pair.

Preflight (A0): 2026-07-10 05:45 โ€” load1 2.33/32c (โ‰ค24) ยท no heavy jobs ยท 38 GiB avail, si trickle/so=0 ยท CH + sidecar + kintsugi active โ†’ ALL PASS. Sync: origin/main unchanged. Provenance: causalnex 0.12.1 Apache-2.0 verified; installs only on pyโ‰ค3.10 (numpy<1.24 pin) โ€” maintenance risk recorded, ran via uv py3.10. One capped run, 0.42 min, 18/18 cells, readonly=2 loader, real rows only.

Exam results (pre-registered form, dynotears_exam.py โ€” bounds DERIVED from the known truth, not tuned)

LegReadingRuling
E1 โ€” the certain coefficient (true ฮฒ = 1.0 exactly)Top twin edge correctly identified in all 18 runs โ€” but its weight reads 0.97 (ฮป=0.05) โ†’ 0.94 (ฮป=0.1) โ†’ 0.556 (ฮป=0.2): a 3%โ†’44% L1-shrinkage bias on the ONE known value, load-bearing and underivedFAIL 12/18
E2 โ€” no phantom structure on the twinUp to 13 spurious twin edges with |w| up to 0.29 (bound: 0.1) in 9/18 runs โ€” the L1-collinearity pathology: an exact duplicate's signal is split across imaginary lagged edges. The tigramite siblings drew exactly ONE edge, 12/12 each (rows 57, 65)FAIL 9/18
F014 trap0 crashes, 0 NaN weights โ€” it fails by being WRONG, not blind: exactly what the entrance exam exists to catch, at the cheapest priceCLEAN (and damning)
Verdict (row 67)EXCLUDE(ground-truth exam failed โ€” ฮป-load-bearing bias 3%โ†’44% on the known coefficient; phantom twin edges to |w|=0.29 in 9/18 runs) โ†’ BONEYARD; re-proposal only with debiased/adaptive penalties + a collinearity-safe formulation
STRUCTURE-LEARNING ON THIS SUBSTRATE โ€” 0 FOR 3:

  PCMCI+     (constraint-based)         exam 12/12 โœ“ โ†’ census ฮฑ-fragile 0.625  โœ— (row 58)
  LPCMCI     (constraint + latent)      exam 12/12 โœ“ โ†’ census ฮฑ-fragile 0.393  โœ— (row 66)
  DYNOTEARS  (continuous optimization)  exam FAIL: ฮป-biased truth + phantoms   โœ— (row 67)

  three schools, three kills, three DIFFERENT failure modes โ€” all on record

โ–ถ Next iteration

Iteration 56 ยท 2026-07-10 ยท EXCLUDE (row 67, BONEYARD +1) โ€” kills are wins, cheap kills doubly so ยท capped (0.42 min) ยท readonly=2 ยท zero generated values ยท append-only ยท evidence: dynotears_exam.py ยท dynotears_exam_results.json