Blame for protocol_lock.json

{
  "code_sha256": "8ad2c436ed95891f75738c067be486b53f85276e605ee7581fbb65e64c15a72a",
  "locked_utc": "2026-11-06T08:25:53.345402+00:00",
  "development_seeds": [
    1,
    2,
    3
  ],
  "holdout_seeds": [
    101,
    102,
    103,
    104,
    105,
    106,
    107,
    108,
    109,
    110,
    111,
    112,
    113,
    114,
    115,
    116,
    117,
    118,
    119,
    120
  ],
  "group": "D_16, 32 elements",
  "conditions": [
    "structured",
    "asymmetric",
    "constant"
  ],
  "budgets": [
    250000,
    2000000
  ],
  "per_seed_train_tasks": 12,
  "per_seed_fresh_tasks": 24,
  "fresh_lengths": [
    4,
    8,
    16,
    32
  ],
  "fresh_per_length": 6,
  "primary_comparison": "recursive vs fast, with identical full-word dictionary and segmentation; dense and static isolate emergent symmetry",
  "primary_metric": "all algorithm work including training, discovery, parsing, materialization, validation and dictionary transactions",
  "fixed_budget_metric": "fresh tasks exactly correct within total 250000 algorithm-operation budget",
  "reduction_control": "symmetry_reference deliberately identical to recursive, so no advantage over symmetry-aware memoized DP is claimed",
  "negative_controls": "asymmetric inputs and constant kernels; all results reported regardless of sign",
  "causal_probe": "h1 to h2 to h3 to h4 in subgroup ladder; actual boundary kernels change; learned, erased, inherited, static arms",
  "oracle": "independent sequential per-target DP; fixed before methods, cost reported separately and added equally",
  "wall_clock": "single instrumented runs per seed; descriptive Python timing, not production performance",
  "memory": "retained DAG scalar slots include uncached ancestors; not bytes or peak temporary memory",
  "no_tuning": "Do not edit code or parameters after viewing held-out results except disclosed correctness fixes and a new holdout block."
}