Iteration 22 — The Entrance Exam Falsified Itself: STABLE Is an ε-Frontier

Iteration 22 — The Entrance Exam Falsified Itself: STABLE Is an ε-Frontier EXAM v1 KILLED CANDIDATE A ALIVE

In plain words: the new e-value instrument sat the entrance exam — identify the one relationship we "knew" was perfectly regime-invariant (the committed duplicate pair) — and rejected it, loudly (e=43). Before blaming the instrument, the loop measured the ground truth itself: the "perfect" duplicate is actually 99.9999% linear, not 100%. The remaining dust — two millionths of the variance — is itself regime-structured, and with twenty thousand data points any honest test must see it. So the exam's premise was wrong, not the instrument: no valid test can pass an exactness exam on real data at our sample sizes. The deep consequence lands on the campaign's one open goal: "perfectly stable" is not a yes/no question on real markets — it is a tolerance question ("stable up to how much dust?"), and the tolerance must be written down before testing. Even the old instrument's behavior gets partially rehabilitated in hindsight — it too may have been seeing real dust — though its rejection stands, because it has no tolerance dial at all. The new instrument does everything right (its calibration and its zero-guard both verified live) and proceeds to the re-specified, tolerance-anchored exam next firing.

Preflight (A0): 2026-07-08 09:19 UTC — load1 5.30/32c (≤24) · 22 GiB avail · CH + sidecar active → ALL PASS. Rebased on moved main (5 commits). Two capped runs (0.21 min exam + measurement probe), readonly loaders, permutations of real rows only.

EXAM v1 (exactness form): identify {ofi} for target = turnover_imbalance e({ofi}) = 43.4 → REJECTED, no degenerate-zero flag → MEASURE THE GROUND TRUTH: pooled Pearson 0.999999 · per-env Spearman 0.999967–0.999999 → ε-exact, NOT exact; residual dust ~2e-6 of variance, env-structured → at n≈20k any calibrated test MUST reject → KILLED-QUESTION (row 29): exact-invariance identification is unsatisfiable as an exam or a goal at our N — STABLE is formally an ε-FRONTIER → row 30 corrects iter-12's mechanism attribution (verdict unchanged: causalicp has no tolerance dial → blind for the ε-question) → CANDIDATE A ALIVE: calibrated e-values ✓ · degenerate guard verified ✓ NEXT: ε-relaxed exam — tolerance ε₀ anchored to the committed duplicate's own residual heterogeneity (real-data-derived, no fiat); identified = subsets indistinguishable from duplicate-level dust.

▶ Next iteration

Next firing: pre-register the ε-relaxed entrance exam (ε₀ = the duplicate pair's own measured residual-heterogeneity scale — committed-anchored, tagged DERIVED) and run it: candidate A must accept {ofi} at ε₀ and reject non-ofi subsets at the same ε₀. Pass → ladder walk. Standing for the operator: the 73 cycle-1 draft statuses and the Frontier #4 precondition review.

Iteration 22 · 2026-07-08 · exam v1 killed by measurement (rows 29–30) · candidate A alive · capped · zero synthetic data · append-only