In plain words: after five exam forms across three firings, the new instrument finally did the thing no prior instrument could: handed real market data containing one genuinely perfect regime-invariant relationship (the byte-identical duration twins โ re-verified today: zero unequal rows in all ten regimes), it accepted every answer containing the truth, rejected every answer without it, and identified exactly the right variable. Two last obstacles fell this firing: the power gap (fixed by multiplying independent per-fold evidence โ v4) and a pure numerics gremlin (raw microsecond scales made the math itself hallucinate regime-structure at the 18th decimal; standardizing the columns silenced it โ v5). The instrument's zero-guard fired perfectly โ five folds, five exact zeros. The exam saga also leaves behind hard campaign knowledge: exact stability is a tolerance question on real data; the crypto "duplicate" is regime-structured grit; and every future regression harness must standardize its columns. Candidate A now enters the formal six-gate ladder.
Preflight (A0): 2026-07-08 11:18 UTC โ load1 3.78/32c (โค24) ยท 22 GiB avail ยท CH + sidecar active โ ALL PASS. Three capped runs (v4 0.52 + identity probe + v5 0.5 min), readonly=2, permutations of real rows only.
v4 complete-case + fold-product (J=5, K=99): E2 FIXED (wrong subsets e up to 1147)
but E1 broke, 0 degenerate folds โ identity probe: twins EQUAL in all 10 envs โ
diagnosis: cond(X'X) ~ 1e18 (raw microsecond scales) โ float64 cancellation
masquerading as heterogeneity, amplified by the product
v5 + globally standardized anchor columns (deterministic; residuals unchanged
in exact arithmetic):
E1 โ all duration-subsets e_total = 1.0, degenerate folds 5/5 (exact zeros)
E2 โ all wrong subsets rejected (qps+burstiness e = 960)
IDENTIFIED = {duration_us} โ EXAM PASS (LEDGER row 33)
โถ Next iteration
Next firing: candidate A's formal ladder walk. M3 redundancy attack vs the admitted stack (what failure mode does e-ICP-fp bring that k_v/G0โG11 don't โ with the STABLE identification role as the obvious answer to defend), then M4/M5โฒ on the persistence panel. ADMIT would fill the stack's one vacant instrument slot. Then H-016 seqICP (B-01 continues). Standing for the operator: the 73 cycle-1 draft statuses and the Frontier #4 precondition review.
Iteration 24 ยท 2026-07-08 ยท entrance exam PASSED (row 33) ยท capped ยท zero synthetic data ยท append-only