Iteration 23 — Exams v2 + v3: True Anchor Found, a Power Gap Remains CHECKPOINT

In plain words: two more rounds of the entrance exam, two more honest lessons. Round two tried the tolerance idea ("accept anything as regime-stable as the duplicate's own dust") and failed backwards: wrong answers have proportionally smoother residuals than the duplicate's dust, so the tolerance axis doesn't rank correctness at all — and with that, the crypto duplicate is retired as a test case (its 0.0001% imperfection is regime-structured rounding grit). But the campaign owns a genuinely perfect case: the forex duration twins, byte-for-byte identical across 701,424 rows. Round three ran on that true anchor: the instrument never rejects the truth (every duration-containing answer accepted — validity confirmed), but it also fails to reject the wrong answers (they read as "mildly stable" and slip under the bar) — a power shortfall of the simple e-value formula used, not a validity flaw. The candidate stays alive: the source paper provides exactly the stronger (optimal) constructions for this, and that is the next slice.