{
  "date": "2026-09-07",
  "reason": "Before experiment captures: the original 1% variable-length residual tolerance failed (1.442%). A separate B diagnostic gives zero difference for changed future tokens at identical sequence length and exact repeat equality. Length changes move the primary J ratio by 0.00450/0.00193 and output ratio by -0.06451/+0.04116 on this diagnostic. Replace the conflated gate with strict same-shape causal and repeat tests (relative L2 <1e-6), retain length sensitivity per arm. No trait data inspected; exact-prefix primary scans unchanged.",
  "initial_commit": "d96c69c",
  "approved_source_sha256": {
    "probes/express01.py": "bd744e154e563465657c9cfde06ab4be12e03ca24413bcf5ef669d73b8406784"
  },
  "analysis_sha256": "f673df809492b925441e2d89bb52c7eb622cea665f6a730c409eb35fec950990",
  "diagnostic_exit": 0
}