โ€บNavigation

Iteration 27 โ€” H-016 seqICP Passes the Entrance Exam EXAM PASS ยท LADDER PENDING

In plain words: the candidate queue resumed with seqICP โ€” an invariance test with one genuinely new trick: it needs no regime labels at all. Where every instrument so far had to be told "these are the ten regimes," seqICP slices time into blocks by itself and asks whether a relationship holds steady across them. It faced the same ground-truth exam that took the previous candidate five attempts: real data containing one perfectly invariant relationship (two byte-identical columns). It passed on the first form โ€” accepted every answer containing the truth, rejected every answer without it, identified exactly the right variable, in under 30 seconds. Two honesty notes: the shuffle-based rejection check ignores that neighbouring bars resemble each other (could make rejections too easy โ€” flagged, same note as the previous candidate carries), and the exam pinned one block layout where the full method scans many. Next comes the hard part: the certificate slot this candidate targets was just filled โ€” so its cheapest kill is now the redundancy gate: does it measure anything the newly admitted identifier doesn't?

Preflight (A0): 2026-07-09 โ€” load1 3.34/32c (โ‰ค24) ยท 23 GiB avail ยท si/so โ‰ˆ 0 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. Worktree rebased on main (2 docs commits) before work.

Exam design (pre-registered in-script before running)

ElementChoice
Candidate mechanismseqICP (Pfisterโ€“Bรผhlmannโ€“Peters 2018 JASA): pooled OLS per subset; environments constructed from the TIME structure โ€” B=10 contiguous blocks over the pooled chronological series, zero env labels; decoupled block-mean + block-variance residual statistics; Bonferroni combination; K=99 permutation null
Data conventions (v5, inherited)10 EURUSD@5 envs pooled ยท complete-case rows ยท globally standardized design + target ยท degenerate zero-residual guard ยท target intra_duration_us = byte-identical twin of anchor duration_us
PASS criteriaE1 every duration-containing subset accepted ยท E2 every other subset rejected (ฮฑ=0.05) ยท identified intersection = {duration_us}

Result

SubsetpRead
duration_us (and all 3 supersets)1.0 (degenerate โ€” exact zero residuals)ACCEPTED โœ“
intra_intensity_qps0.02 (T_mean 12.6)REJECTED โœ“
intra_burstiness0.02 (T_mean 12.8)REJECTED โœ“
intensity+burstiness0.02 (T_mean 11.9)REJECTED โœ“
Identified set{duration_us} โ€” exactly right. E1 โœ“ ยท E2 โœ“ ยท EXAM PASS (row 36)

Recorded caveats (verify-before-report): (i) row-permutation null breaks serial dependence โ†’ E2 possibly anticonservative; block-null refinement is M2 material (the same standing note e-ICP-fp carries from row 34); (ii) the bounded form fixes one block decomposition where the paper scans a grid. Neither touches E1 (driven by exact zero residuals, null-free).

The looming gate: seqICP targets the STABLE-identification role โ€” which rows 34โ€“35 just filled on both substrates with e-ICP-fp. M3's redundancy attack against e-ICP-fp's per-feature outputs is therefore the cheapest kill: |ฯ| โ‰ฅ 0.95 โ†’ EXCLUDE(redundant). Survival requires either low correlation or a demonstrably uncovered failure mode โ€” the candidate's claim is env-free operation (works where regime labels are wrong or absent).

โ–ถ Next iteration

Iteration 27 ยท 2026-07-09 ยท CHECKPOINT (row 36) โ€” exam PASS on first form ยท capped ยท zero synthetic data ยท append-only ยท evidence: seqicp_exam.py ยท seqicp_exam_results.json