โ€บNavigation

Iteration 12 โ€” F014 Falsified as BLIND-GAUGE ยท The STABLE Ceiling Re-Opens KILL #3 ยท ONE-SHOT RESOLVED

What happened in this iteration, in plain words: the campaign's long-standing hope was that a smarter (nonlinear) version of the invariance instrument would settle whether any feature can hold its behavior in every regime. Before trusting it, the rulebook demands the power test: hand the instrument a case whose answer we know with certainty. We have one made of pure real data โ€” a feature that is a byte-for-byte committed duplicate of one of the core signals, so its relationship to that signal is perfectly invariant everywhere. Any working identification procedure must nail it. Both versions failed. The linear one rejected everything โ€” including the exactly-true answer (perfect fits blow up its math); the nonlinear one accepted eight candidate answers that don't overlap, identifying nothing. Conclusion with teeth: this instrument family cannot identify invariance even when invariance is perfect โ€” so every "nothing is invariant" it ever produced (including the campaign's famous 0-of-91) was blindness, not evidence. The falsifier is falsified; the STABLE question is fully open again, and what's missing is not data but a better instrument. Third kill of the campaign โ€” and the most consequential one.

Preflight (A0)

2026-07-03 15:18 UTC โ€” load1 3.23 / 32 cores (โ‰ค24) ยท no competing nasimubd jobs ยท 29 GiB available ยท clickhouse-server active ยท sidecar healthy โ†’ ALL PASS. One capped readonly run, 0.19 min, zero generated values (the ground-truth case is a committed real-data duplicate โ€” no synthetic plants).

The power test and the double failure

GROUND TRUTH (real data, committed): turnover_imbalance โ‰ก ofi (|ฯ|=1.0, all cells)
  โ†’ the relationship turnover_imbalance ~ ofi is PERFECTLY invariant in all 10 envs
  โ†’ any identification procedure MUST return an estimate containing {ofi}

  linear ICP (v3 wiring):  n_accepted = 0  โ†’ REJECTS ALL, incl. the true set
                            (zero-residual degeneracy explodes the test)
  rank ICP (monotone):     n_accepted = 8, intersection = โˆ… โ†’ identifies NOTHING

  โ†’ BOTH variants: M5' power calibration FAIL on a perfect case
  โ†’ EXCLUDE(BLIND-GAUGE) โ€” LEDGER row 17, BONEYARD (revisit path recorded)

THE CORRECTION CHAIN OF "ICP 0/91", now three layers deep:
  committed reading : "0/91 features regime-invariant"           (2026-06)
  iter 7            : not rejection โ€” non-identification          (accepted
                      families non-empty for all 100 features)
  iter 12           : non-identification is NOT EVIDENCE either โ€” the
                      identifier fails a perfect ground-truth case
  โ†’ the STABLE ceiling is an INSTRUMENT GAP, not a settled fact.
    Missing: an identification-capable invariance framework (handles exact/
    degenerate relationships, reports confidence sets, not intersections).
    Pragmatic STABLE evidence meanwhile: the k_v = 10/10 flags (iter 10/11)
    + expiry re-grounding every completed regime slice.

Technical record

ItemValue
Ledger / Boneyardrow 17 โ€” EXCLUDE(BLIND-GAUGE) at M5โ€ฒ; F014's 2026-06-16 INCLUDE-IF one-shot RESOLVED; boneyard entry with revisit path (identification-capable framework from a fresh SOTA sweep)
Frontieritem 3 CLOSED (resolved by falsifying the falsifier); Tier-3 refill: SOTA sweep for identification-capable invariance candidates
Real-subject readout (void by blindness, recorded)0/26 identified in both variants โ€” carries no evidence per the M5 doctrine
Numberslinear gt: n_accepted 0 ยท rank gt: n_accepted 8, estimate โˆ… ยท wall 0.19 min ยท ฮฑ=0.05 FROZEN ยท rank transform parameterless (DERIVED)
Harnessf014_identifiability_falsifier.py ยท f014_results.json
Compute / prod impactone 0.19-min capped readonly run ยท zero writes outside audit folder + dashboard

โ–ถ Next iteration

Iteration 12 ยท 2026-07-03 ยท Frontier #3 resolved: falsifier falsified (kill #3) ยท capped read-only ยท zero synthetic data ยท append-only