{ "schema": "T_GPT_PROTOCOL_V0_8_2", "date": "2026-09-11", "status": "STATISTICAL_REPAIR / NO_CONFIRMATORY_COLLECTION_BEFORE_FREEZE", "supersedes_for_current_design": "protocol-v0.8.1.json", "preserves_history": true, "research_position": "instrumentation-first; T^FIELD is a candidate method that can be refuted or retired", "metric_spec": "/T-GPT/data/metrics-v0.8.2.json", "review_trigger": "/T-GPT/revues/claude-02/", "confirmatory_gate": { "status": "CLOSED", "opens_only_if": [ "all v0.8.2 self-tests and adversarial tests pass", "freeze manifest completed with hashes and immutable IDs", "cluster assignment alpha >= 0.67 and cluster relevance alpha >= 0.67 on development audit", "task-count/precision plan satisfied", "neutral-probe and paired-fork implementation verified", "condition-blinding audit completed", "all primary endpoints and contrasts reduced to one per experiment" ] }, "phases": { "planning_tasks": { "purpose": "estimate between-task variance and debug task families; never enter confirmatory effect estimates", "minimum_tasks": 12 }, "task_specific_development": { "purpose": "for every future confirmatory task, build a balanced condition-blind semantic codebook using independent development generations; those generations are excluded from confirmation", "minimum_generations_per_condition": 48, "recommended": 64 }, "freeze": { "required": [ "task registry and family labels", "frozen semantic codebook per task", "cluster-level relevance adjudication", "assignment rules and normalizer version", "probe texts and neutral control", "model/checkpoint IDs", "sampling parameters", "input/output token budgets", "primary endpoint and single primary contrast per experiment", "analysis script SHA-256/Git SHA", "power/precision report", "annotation agreement report", "blindness audit plan" ], "change_after_freeze": "requires new freeze_id and invalidates the previous confirmatory preregistration for new data" }, "confirmation": { "task_unit": true, "minimum_tasks": 30, "target_tasks": 40, "final_task_count_rule": "determined before confirmation by independent planning-task variance/precision simulation; if >60 tasks would be required for the preregistered precision target, redesign rather than silently underpower", "per_task_generations": "determined by precision simulation and budget; N=64 is a starting point, not a sacred constant" }, "replication": { "tasks": "new independent tasks/families where feasible", "purpose": "replicate any claim promoted beyond local task-family scope" } }, "task_precision": { "primary_goal": "task-level 95% interval half-width target <= 0.07 for the family-level primary effect unless planning data justify a stricter target", "minimum_power": 0.80, "alpha": 0.05, "rule": "report both precision-based and effect-size-based planning; choose the larger preregistered task count before confirmation" }, "mapping_and_annotation": { "final_assignment_states": ["FROZEN_CLUSTER_ID", "UNMAPPED", "UNRESOLVED"], "uncertain_must_be_adjudicated": true, "cluster_relevance_is_cluster_level": true, "output_assignment_is_output_level": true, "aggregate_representation_is_separate_annotation_task": true, "agreement_gate": { "minimum_alpha": 0.67, "target_alpha": 0.80, "failure_action": "return to development; revise codebook; issue new freeze_id; do not run primary analysis" }, "author_primary_reference_annotator": false, "condition_labels_removed": true, "T_vocabulary_forbidden": true, "blindness_measurement": "annotators guess condition on a preregistered audit subset; guess accuracy is reported" }, "budget": { "primary_matching": "total input + output tokens", "mandatory_secondary": ["input_tokens", "output_tokens", "max_output_tokens", "wall_clock_latency", "serial_depth", "parallelizable_calls", "estimated_monetary_cost"], "E3_identical_output_cap": true, "E3_identical_frozen_pool": true }, "experiments": [ { "id": "E1_STAGE", "priority": 1, "question": "At which post-training transition does directional semantic distribution shift occur after null calibration?", "preferred_trajectory": ["BASE", "SFT", "DPO", "RLVR"], "template_ablation": true, "primary_contrast": "one preregistered end-to-end or single adjacent contrast selected before confirmation from planning rationale", "primary_endpoint": "DEFICIT_EXCESS", "always_report_with": ["EXPANSION_EXCESS", "Q_task", "Q_fact", "mapping_rates", "cost"], "transition_decomposition": "secondary; Holm-adjusted within family", "null_calibration": "label permutation preserving group sizes and task", "minimum_model_families": 2 }, { "id": "E2_RECOVERY", "priority": 2, "question": "Does an alternatives-oriented recovery probe move semantic mass toward the reference distribution more than an equal-cost neutral recheck from the same contracted generation?", "primary_endpoint": "RECOVERY_LIFT", "primary_contrast": "RECOVERY_PROBE vs NEUTRAL_RECHECK", "reference_null": "independent A_prime replicate preferred; permutation calibration also reported", "paired_forks": "each B_i forks into RECOVERY and CONTROL continuations from the same starting output/context", "paired_randomization": true, "budget": "A, B+RECOVERY and B+CONTROL comparisons report and match total token budgets as preregistered", "retired_primary_metrics": ["REC_GAIN", "IRREV_OBS"] }, { "id": "E3_TOPOLOGY", "priority": 3, "question": "Which topology/aggregation retains more relevant frozen-pool coverage under identical pool and output budget without increasing false representation?", "primary_endpoint": "AGG_COVERAGE", "primary_contrast": "T_WEAVE vs strongest preregistered non-T baseline", "mandatory_guardrail": "REPRESENTATION_FPR", "conditions": ["STANDARD_SUMMARY", "MAJORITY_OR_JUDGE", "EXHAUSTIVE_NON_T", "COVERAGE_BASELINE_V0_8_2", "T_WEAVE"], "representation_annotation": "true pool clusters plus preregistered distractor clusters mixed blindly", "output_normalization": "fixed proposition-list representation before annotation; raw aggregate retained", "identical_output_token_cap": true }, { "id": "E4_R_META", "priority": 4, "question": "Does structural recurrence exceed mirroring/elicitation under matched non-structural controls?", "primary_endpoint": "STRUCTURAL_RECURRENCE_RATE_DIFFERENCE", "primary_contrast": "STRUCTURAL_CRITIQUE vs pooled STYLE_CONTROL + LENGTH_CONTROL + NO_CRITIQUE", "arms": ["STRUCTURAL_CRITIQUE", "STYLE_CONTROL", "LENGTH_CONTROL", "NO_CRITIQUE"], "blinded_coding": true, "counter_hypothesis": "CH11_MIRROR_ELICITATION" } ], "strong_baselines": [ "DIRECT", "EXPLICIT_K_DISTINCT", "MINIMAL_MULTI", "DISPERSIVE_SAMPLING", "VERBALIZED_SAMPLING", "COVERAGE_BASELINE_V0_8_2", "EXHAUSTIVE_NON_T", "BEST_OF_N_OR_CALIBRATED_JUDGE", "STANDARD_MULTI_AGENT_DEBATE" ], "analysis": { "within_task": "design-matched permutation/randomization; no v0.8.1 threshold-bootstrap", "across_tasks": "task-level inference only for generalized claims", "one_primary_endpoint_and_contrast_per_experiment": true, "secondary_correction": "Holm within experiment family", "null_and_negative_results_publishable": true, "no_mean_of_task_ratios": true }, "auto_null": { "T_layer_removed_or_retired_if": [ "strongest non-T baseline equals/exceeds T method within uncertainty at lower/equal cost", "T_signature appears without coverage/performance benefit", "effect disappears under blinded replication", "mapping/blinding quality is insufficient for the claimed interpretation" ] } }