{
  "schema": "T_GPT_PLANNING_PROTOCOL_V0_8_2",
  "date": "2026-09-11",
  "status": "PLANNING_GATE_OPEN_CONFIRMATORY_GATE_CLOSED",
  "task_registry": "/T-GPT/data/planning-tasks-v0.8.2.json",
  "principle": "Planning estimates design parameters only. No planning task, output, codebook or effect estimate may enter confirmatory inference.",
  "P0_task_qualification": {
    "purpose": "Verify that tasks produce non-trivial, adjudicable semantic distributions rather than ceiling/single-class outputs.",
    "minimum_generations_per_task_condition": 64,
    "condition": "neutral independent sampling condition selected before P0 execution; exact model, checkpoint and sampling recorded in a planning run manifest",
    "codebook": "condition-blind development codebook built only from P0 outputs",
    "qualification_gates": {
      "minimum_relevant_clusters": 6,
      "maximum_relevant_clusters_for_primary_manual_codebook": 24,
      "minimum_mapped_plus_adjudicated_rate": 0.80,
      "maximum_UNRESOLVED_rate": 0.10,
      "cluster_assignment_alpha_minimum": 0.67,
      "cluster_relevance_alpha_minimum": 0.67,
      "top_cluster_mass_maximum": 0.80,
      "minimum_clusters_with_empirical_mass_ge_0_03": 4
    },
    "failure_rule": "A task failing any hard gate is not silently repaired into qualification. One preregistered revision of wording/codebook is allowed and creates a new planning_task_version. Original failure remains logged.",
    "minimum_qualified_tasks_to_continue": 12,
    "family_rule": "At least 2 qualified tasks from each of the five families are required; otherwise redesign the family before P1."
  },
  "P1_effect_planning": {
    "purpose": "Estimate between-task variance and precision requirements for the actual v0.8.2 endpoints without making scientific claims.",
    "tasks": "only P0-qualified planning tasks; still excluded from confirmation",
    "E1": {
      "run_only_after": "candidate primary model trajectory and one primary contrast are chosen",
      "estimate": "task-level DEFICIT_EXCESS distribution for the candidate primary contrast"
    },
    "E2": {
      "run_only_after": "RECOVERY_PROBE and NEUTRAL_RECHECK texts, paired-fork mechanism and token caps are fixed for planning",
      "estimate": "task-level RECOVERY_LIFT distribution with paired B_i forks"
    },
    "E3": {
      "run_only_after": "aggregate proposition normalizer and distractor construction are fixed for planning",
      "estimate": "paired task-level AGG_COVERAGE difference and REPRESENTATION_FPR between candidate strongest non-T baseline and T_WEAVE"
    },
    "E4": {
      "status": "deferred until E1-E3 planning gates pass",
      "estimate": "task-level structural recurrence contrast only on a separate planning subset or newly generated planning tasks"
    }
  },
  "task_count_decision": {
    "tool": "/T-GPT/tools/plan_task_precision_v0.8.2.py",
    "minimum_confirmatory_tasks": 30,
    "initial_target": 40,
    "maximum_without_redesign": 60,
    "precision_target_half_width_95": 0.07,
    "power_target": 0.80,
    "alpha": 0.05,
    "minimum_effect_of_interest": "must be justified before reading confirmatory data; planning estimates may inform feasibility but not redefine it post hoc",
    "decision_rule": "freeze task count = max(30, precision_required_n, power_required_n); if >60, redesign or NO-GO"
  },
  "confirmatory_task_generation": {
    "starts_only_after": "P1 variance report is complete",
    "independence": "new prompts/scenarios not used in P0 or P1",
    "family_balance": "five families retained unless P0/P1 demonstrate a preregistered family is non-adjudicable; no family >40% of final tasks",
    "registration": "all confirmatory tasks receive immutable task_id, family_id, prompt hash and version before any confirmatory model call",
    "task_specific_development": "codebooks for confirmatory tasks are built from separate development generations, never from confirmatory outputs"
  },
  "freeze_entry_gates": [
    "planning registry validation passes",
    "P0 has >=12 qualified tasks and >=2 per family",
    "P1 variance/precision report exists for the primary experiment to be frozen",
    "minimum effect of interest is declared",
    "confirmatory task count is fixed before confirmatory calls",
    "confirmatory task registry is complete and distinct from planning registry",
    "all task-specific codebooks/relevance decisions satisfy alpha >=0.67",
    "blindness audit plan is executable",
    "v0.8.2 adversarial CI is green",
    "freeze manifest hashes/models/sampling/budgets/endpoints are complete"
  ],
  "current_gate_state": {
    "planning_gate": "OPEN",
    "P0_executed": false,
    "P1_executed": false,
    "confirmatory_task_registry_created": false,
    "freeze_manifest_completed": false,
    "confirmatory_gate": "CLOSED"
  }
}