{
  "proposer_model": "gpt-6-astra",
  "status": "Six exact schedules and two-stage selection frozen before new candidate development feedback",
  "identities": {
    "origin": {
      "repo_id": "HuggingFaceTB/SmolLM2-360M-Instruct",
      "revision": "a10cc1512eabd3dde888204e902eca88bddb4951"
    },
    "donor": {
      "repo_id": "HuggingFaceTB/SmolLM2-360M",
      "revision": "f8027fd0eaeea54caa13c31d31b9fdc459c38b49"
    }
  },
  "candidate_ids": [
    "L1",
    "L2",
    "L3",
    "L4",
    "L5",
    "L6"
  ],
  "adaptive_schedules": false,
  "gates": {
    "arc_minimum_delta_lower_bound": 0.02,
    "instruction_minimum_delta_lower_bound": -0.05,
    "bootstrap_draws": 4096,
    "family_alpha": 0.05,
    "family_metrics": 2,
    "baseline": "original instruct",
    "scorer": "Unchanged existing scorer, prompts, tokenizer, precision, decoding and strict Google instruction checks."
  },
  "screen": {
    "arc_examples": 1280,
    "instruction_examples": 256,
    "data": "Identical original blend-search development set; all previously exposed.",
    "candidates": "All six frozen schedules",
    "advance_count": 2,
    "ranking": {
      "eligible_first": true,
      "eligible_order": [
        "higher ARC accuracy",
        "higher instruction accuracy",
        "lower stored-parameter-weighted mean donor fraction",
        "lexicographically smaller candidate ID"
      ],
      "ineligible_order": [
        "higher min(ARC_lower_bound - 0.02, instruction_lower_bound + 0.05)",
        "higher ARC accuracy",
        "higher instruction accuracy",
        "lower stored-parameter-weighted mean donor fraction",
        "lexicographically smaller candidate ID"
      ],
      "units": "Accuracy fractions; gate lower bounds use the configured two-gate family correction. Parameter-weighted mean includes all stored tensors, including zero-alpha and nonlayer tensors, counted once."
    },
    "rule": "Advance exactly the first two candidates under this ranking, even if fewer than two are eligible."
  },
  "development_validation": {
    "arc_examples": 1280,
    "instruction_examples": 512,
    "data": "Most recently opened former holdout, now development validation.",
    "candidates": "Only the two advanced candidates",
    "additional_acceptance_gate": false
  },
  "final_development_selection": {
    "arc_examples": 2560,
    "instruction_examples": 768,
    "data": "Pool screen and development-validation per-case outcomes; recompute paired-bootstrap gates on the pooled data. Do not average stage-level bounds.",
    "candidates": "Only the two advanced candidates",
    "ranking": {
      "eligible_first": true,
      "eligible_order": [
        "higher ARC accuracy",
        "higher instruction accuracy",
        "lower stored-parameter-weighted mean donor fraction",
        "lexicographically smaller candidate ID"
      ],
      "ineligible_order": [
        "higher min(ARC_lower_bound - 0.02, instruction_lower_bound + 0.05)",
        "higher ARC accuracy",
        "higher instruction accuracy",
        "lower stored-parameter-weighted mean donor fraction",
        "lexicographically smaller candidate ID"
      ],
      "units": "Accuracy fractions; gate lower bounds use the configured two-gate family correction. Parameter-weighted mean includes all stored tensors, including zero-alpha and nonlayer tensors, counted once."
    },
    "selected_count": 1,
    "eligibility": "Both pooled development gates must pass.",
    "fallback": "If neither qualifies, select the first under ineligible_order and permanently mark diagnostic_only.",
    "freeze": "Seal selected label, exact checkpoint hash and eligible/diagnostic_only status before any new holdout evaluation.",
    "no_rescue": "No holdout-based edits, reselection, or promotion of a diagnostic fallback."
  },
  "controls": [
    "original instruct",
    "full rollback",
    "fixed uniform 15% base-donor blend B15"
  ],
  "cache": "Reuse exposed observations only for exact checkpoint/scorer/configuration/case identity; no new learning or cache reuse for fresh rows.",
  "fresh_holdout": {
    "sampling_owner": "Parent prepared protocol; availability verification pending. No proposer access to new examples or partial results.",
    "planned_arc": {
      "ARC_Easy_unused_official_train": 640,
      "ARC_Challenge_unused_official_train": 640
    },
    "planned_instruction": {
      "unused_Ai2_same_supported_checkers": 512
    },
    "requirement": "Freeze exact achieved sources, counts, admission, sampling and deduplication before new candidate feedback. Exclude every prior exposed case/task/body and disclose official ARC split and instruction source-mixture changes.",
    "secondary_reporting": "Report source and difficulty strata descriptively; do not use them to rescue eligibility.",
    "outcome": "Evaluate one sealed candidate and all three controls once. A diagnostic fallback cannot be accepted or exported as a repair."
  },
  "limitations": [
    "Two-stage screening may eliminate a candidate that would rank better on pooled development.",
    "Pooled development and fresh holdout have differing source mixtures; neither is an IID enlargement of original IFEval.",
    "Architecture-based schedules are hypotheses, not established localization of knowledge or instruction following."
  ],
  "native_contract": {
    "max_operations": 16,
    "max_exact_names_per_operation": 200,
    "grouping": "Group disjoint exact names by positive alpha; omit zero-alpha targets. All tensor mapping entries remain explicit.",
    "constraints": "No replacement payloads, training, core/scorer edits, or overlapping operations."
  }
}