Iteration 11 — STABLE screen S-03 13 PASS · 1 OUT FEED RE-TRIGGERED (B-03)

In plain words: round two of interrogation for the campaign's priority boundary. Fourteen instruments accumulated since the first STABLE screen (the count itself was off by one in last firing's estimate — corrected on the record). Thirteen passed: the causal-structure machines with their autocorrelation-aware tests, the regularity-free and stopping-safe hypothesis frameworks, the robustness-margin instruments, and the full anchor triad. One went out with a principle attached: Neural Granger causality — its output is a set of connections chosen by stochastic neural-network training, so even a cleverly built null check can't distinguish "the relationship changed" from "the training run wobbled." The screen drew that line explicitly: its deterministic sibling (DYNOTEARS) passed because its optimizer gives the same answer twice, making a bootstrap anchor meaningful. Notes travel with the passers: two robustness cousins the lab should test together, a three-piece anchor toolkit with distinct roles, and honest flags on every constructed anchor. Thirteen passers means the third handoff batch fires next.

 Same idea, two fates — structure-drift detection across regimes:
   H-030 DYNOTEARS       deterministic optimizer  → same data, same answer
                         → block-bootstrap anchor is MEANINGFUL      → PASS (borderline)
   H-029 Neural Granger  stochastic NN training   → same data, different answer
                         → any null is confounded by training noise  → OUT (POLICY)
 The trust-anchor gate is not about elegance — it is about whether
 the instrument can be caught lying. A gauge that answers differently
 on identical evidence cannot be caught.