{
  "schema": "T_GPT_PROTOCOL_V0_8",
  "date": "2026-09-11",
  "status": "preregistered_before_quantitative_campaign",
  "supersedes_for_future_collection": "pilot-v0.7",
  "historical_note": "v0.7 remains published unchanged as a historical preregistration state.",
  "primary_research_question": "Can semantic support absent from a narrower condition be recovered by preregistered generic probes that do not name the missing configurations?",
  "field_types": {
    "FIELD_TOK": "conditional token/continuation distribution when exposed by the model/API",
    "FIELD_SEM": "empirical distribution of semantic configurations obtained by sampling",
    "FIELD_TRAJ": "graph/set of reasoning or agent trajectories and interactions",
    "FIELD_VIS": "configurations exposed in the final visible output"
  },
  "phases": {
    "exploration": {"n_tasks_min": 24, "purpose": "validate code, probe family and frozen codebook; not used for confirmatory claims"},
    "confirmation": {"n_tasks_min": 64, "purpose": "primary inference on unseen tasks/samples"},
    "replication": {"n_tasks_min": 128, "purpose": "replicate effects judged substantively important on a second model family"}
  },
  "budget": {
    "primary_matching_unit": "total_input_plus_output_tokens",
    "secondary_recorded_units": ["number_of_calls", "latency_seconds", "estimated_cost"],
    "rule": "conditions compared for a primary claim must operate under the same preregistered total-token ceiling",
    "overflow": "truncate or stop according to a fixed pre-collection rule; never after qualitative inspection"
  },
  "task_families": {
    "positive_controls": [
      {"id":"POS_SENSE","type":"known_semantic_ambiguity","requirement":"2-5 preregistered valid interpretations"},
      {"id":"POS_MULTI","type":"finite_multi_answer","requirement":"preregistered gold set of multiple valid answers"},
      {"id":"POS_CLASS","type":"synthetic_solution_classes","requirement":"preregistered finite valid solution classes"}
    ],
    "open_ended": ["WHITE", "OTHER_WITH_FROZEN_INDEX_A"],
    "negative_controls": ["MATH", "UNIT"]
  },
  "generation_conditions": [
    "DIRECT",
    "K_DISTINCT_SINGLE_PROMPT",
    "MINIMAL_MULTI",
    "DISPERSIVE_SAMPLING_IF_SUPPORTED",
    "VERBALIZED_SAMPLING",
    "T_FIELD_GENERATION"
  ],
  "aggregation_conditions_same_frozen_pool": [
    "STANDARD_SUMMARY",
    "MAJORITY_OR_CALIBRATED_JUDGE",
    "COVERAGE_SELECTION_NON_T",
    "EXHAUSTIVE_NON_T",
    "STANDARD_MULTI_AGENT_DEBATE",
    "T_WEAVE"
  ],
  "primary_baseline": "MINIMAL_MULTI + semantic clustering + coverage selection",
  "primary_endpoint": {
    "name": "REC_GAIN",
    "definition": "fraction of semantic clusters present in the broad reference support and absent in the narrow condition that reappear under the preregistered generic probe family",
    "unit": "task-level cluster recovery proportion",
    "inference": "report mean/median as appropriate with 95% bootstrap confidence interval over tasks; task is the primary resampling unit"
  },
  "core_metrics": {
    "ES_LOSS": "observable support in broad reference condition absent from narrow condition, under the frozen semantic-equivalence rule",
    "REC_GAIN": "recovered lost support / lost support under preregistered probes",
    "IRREV_OBS": "1 - REC_GAIN within this protocol, probe family, model state and budget; never weight-level irreversibility",
    "COV_GOLD": "recall of preregistered valid interpretations/answers on positive controls",
    "H_SEM": "semantic entropy using a specified semantic-equivalence method",
    "PROV_SURVIVE": "fraction of source branches/ideas whose provenance remains traceable after aggregation",
    "R_META_RATE": "rate of preregistered structural recurrence classes after critique, compared with elicitation controls",
    "T_SIGNATURE": "rate of protocol-induced recurring T^ vocabulary or structure without task-specific gain",
    "TOKENS": "total input + output tokens"
  },
  "secondary_metrics": ["D_sem", "P_human", "Q_task", "Q_fact", "latency", "cost"],
  "probe_family": {
    "frozen_before_confirmation": true,
    "must_not_name_missing_cluster": true,
    "classes": ["generic_reconsideration", "alternative_interpretations", "assumption_check", "coverage_check"],
    "sensitivity_analysis": "report REC_GAIN separately by probe class and for the union"
  },
  "semantic_equivalence": {
    "primary": "frozen bidirectional entailment/equivalence procedure validated on exploration data",
    "secondary": ["embedding clustering", "threshold sensitivity", "human blinded validation", "bootstrap cluster stability"],
    "rule": "no threshold or clustering algorithm may be selected after seeing confirmatory condition labels"
  },
  "stage_experiment": {
    "id": "E1_STAGE",
    "target_chain": "BASE -> SFT -> DPO -> RLVR when public checkpoints or reproducible recipes permit",
    "template_ablation": true,
    "preferred_open_artifacts": ["Tulu 3 recipes", "OLMo 2 / related open checkpoints where available"]
  },
  "topology_experiment": {
    "id": "E3_TOPOLOGY",
    "claim_already_documented": "communication topology can alter convergence/diversity in some multi-agent settings",
    "residual_question": "which topology and contact timing best preserve relevant coverage and provenance at matched token budget?"
  },
  "r_meta_experiment": {
    "id": "E4_R_META",
    "critique_source": "fixed third-party template or independent model",
    "control_arm": "critique style/length rather than target structure",
    "counter_hypothesis": "CH11_MIRROR_ELICITATION"
  },
  "annotation": {
    "condition_blinding": true,
    "project_author_primary_annotator": false,
    "independent_annotators_min": 2,
    "adjudicator_required_for_disagreement": true,
    "agreement_metric": "Krippendorff_alpha_or_predeclared_equivalent",
    "codebook_frozen_before_confirmation": true,
    "t_vocabulary_forbidden_in_annotation_instructions": true
  },
  "epistemic_outcomes": ["SUPPORTED", "PARTIAL", "REFUTED", "RETIRED"],
  "auto_null": {
    "preregistered": true,
    "triggers": ["no_gain_over_strong_baseline", "T_signature_without_task_gain", "excess_cost_without_gain", "diversity_increase_with_quality_collapse"]
  },
  "global_score": null,
  "winner_default": null,
  "vote_default": null
}