Iteration 78 — the Stability Pair: DOUBLE KILL KILLS #21 · #22 · THE DOOR LAW'S THIRD FAMILY

In plain words: two candidates entered together — one counts how consistently each feature gets picked when the selection is re-run on reshuffled blocks of data; the other is its smarter sibling, which stops counting swaps among near-identical features as instability. Their controls all passed: the byte-twin was treated identically to its original in every single run, and the sibling's adjustment was caught working on camera (nine real clusters of near-twin features, with an honest surprise on record: the adjustment isn't always calming — it can INCREASE the instability reading when a rarely-picked feature joins a jittery cluster). But both died at the same door that has now killed three families: asked whether their regime-to-regime readings differ more than randomly re-split blocks of the same data would produce, neither cleared the bar. The meta-pattern is now unmistakable — and discriminating: resampling-style regime readings drown in honest block noise on this data, while instruments with real boundary signal (the two distance certificates) cleared the same style of bar at three to fifty times. The pool was verified healthy (124 features) before trusting the verdict; two data-engineering adjustments (named windows, availability screening) are on the record.