{
  "schema_version": 1,
  "study_id": "factorized_resource_substitution_v1",
  "amendment_id": "compute_cutoff_20260718T202757Z",
  "status": "prospective_compute_only_amendment",
  "original_protocol_sha256": "1897bb5a70f87dd9e173486f0a137a57b5be2cb159cf4324b7cd95117d928cd7",
  "decision_utc": "2026-07-18T20:27:57Z",
  "original_runner_started_utc": "2026-07-18T19:14:53Z",
  "fixed_compute_deadline_utc": "2026-07-19T19:14:53Z",
  "reason": {
    "category": "fixed_compute_deadline",
    "detail": "Measured two-GPU throughput projected 28-31 hours for the original training-first 100M pipeline, beyond the fixed one-day allocation. The registered 40M primary endpoint fits while preserving the complete factorial and all three fresh seeds."
  },
  "outcome_access": {
    "new_assay_files_present": 0,
    "new_holdout_files_present": 0,
    "aggregate_results_read": false,
    "leaderboard_results_read_for_amendment": false,
    "decision_uses_training_throughput_only": true
  },
  "decision_snapshot": {
    "remote_path": "/home/cc/factorized_resource_v1/amendment_snapshot_20260718T202757Z",
    "snapshot_utc": "2026-07-18T20:28:24Z",
    "dense_l12_s47_words_seen": 25586123,
    "dense_l12_s47_words_per_second": 5907.69,
    "orthogonal_l12_s47_words_seen": 20531746,
    "orthogonal_l12_s47_words_per_second": 4752.50,
    "files_sha256": {
      "assay_files_at_amendment.txt": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
      "audit.json": "4050106413d3340184f4ebe250181dc2764f3bb5f3b4af77fc2c9642e9ad87e4",
      "items.manifest.json": "776dd706e7611f1e5b6f7eea2944fd0ef2a2cd9729238741611646b7542ba8d1",
      "train_resource_dense_l12_s47.log": "207fea804d469cf2c179cc9e17381e25af81fd2e6d93c54e714dc5b05e52282c",
      "train_resource_orthogonal_l12_s47.log": "9b1c38ed99c3cc48272acf1bdaa7ba6cddee5af1544082c39536b9414f34ddb6"
    }
  },
  "unchanged": [
    "primary_question",
    "primary_interaction",
    "geometry_and_depth_cells",
    "seeds_47_53_59",
    "corpus_and_tokenizer",
    "optimizer_and_100M_learning_rate_schedule",
    "frozen_assay_items_and_hashes",
    "heldout_loss_control",
    "paired_seed_as_independent_unit",
    "no_checkpoint_or_leaderboard_selection"
  ],
  "effective_execution": {
    "terminal_exposure_words": 40000000,
    "schedule_total_words": 100000000,
    "saved_checkpoint_words": [
      1000000,
      2000000,
      3000000,
      4000000,
      5000000,
      6000000,
      7000000,
      8000000,
      9000000,
      10000000,
      20000000,
      30000000,
      40000000
    ],
    "evaluation_words": [1000000, 5000000, 10000000, 20000000, 40000000],
    "embargo_aggregate_until_all_models_sealed": true,
    "required_models": 12,
    "failed_run_rule": "Do not replace a seed or cell. Mark the prospective factorial incomplete if any registered 40M checkpoint is unavailable."
  },
  "unexecuted_secondary_endpoints": [
    "paired-seed interaction at 100M",
    "60M, 80M and 100M competence trajectories",
    "durability or asymptotic equivalence"
  ],
  "precision_gate": {
    "independent_units": 3,
    "smallest_relevant_absolute_interaction": 0.05,
    "seed_interval": "two-sided 95% Student-t interval over the three paired seed interactions",
    "exact_direction_test": "one-sided exact sign test; minimum attainable p-value with three seeds is 0.125",
    "multiplicity": "The 40M equal-weight competence interaction is the sole primary estimand. Component, held-out-loss, AUC, parameter, FLOP and wall-time analyses are descriptive secondary evidence.",
    "interpretation_rules": {
      "positive_large_consistent": "Mean interaction is at least +0.05, all seed interactions are positive, and the 95% seed interval is above zero. This supports reduced depth sensitivity under orthogonal geometry, not general compute substitution.",
      "negative_large_consistent": "Mean interaction is at most -0.05, all seed interactions are negative, and the 95% seed interval is below zero. This supports an increased depth cost under orthogonal geometry.",
      "practically_small_precise": "The entire 95% seed interval and every seed interaction lie within [-0.05,+0.05]. This is descriptive compatibility with a small interaction, not universal equivalence.",
      "imprecise_or_mixed": "Every other pattern is indeterminate and may not be reframed as a positive, negative or null discovery."
    }
  },
  "estimand_boundary": "The factorial estimates sensitivity to reducing Transformer depth under matched linguistic exposure. It does not equate word exposure with FLOPs, parameter count, wall time, human cognitive resources or general computational efficiency.",
  "evidence_roles": {
    "prospective_primary": "The twelve fresh resource models at 40M only.",
    "retrospective_100M_context": "Previously trained factorized tournament ablations may contextualize behavior, late persistence and failure modes but are not pooled with the fresh factorial.",
    "prior_mechanistic_context": "The frozen 100M specificity matrix and layerwise causal-use trajectory motivate the resource question but do not mediate or statistically rescue the 40M interaction."
  },
  "claim_boundary": "Results concern early acquisition under one matched Strict-Small recipe. No result establishes durable asymptotic substitution, human equivalence, universal architectural optimality, leaderboard superiority or equal-compute efficiency."
}