{ "schema": "T_GPT_PROTOCOL_V0_8_1", "date": "2026-09-11", "status": "PREREGISTRATION_FRAMEWORK_BEFORE_CONFIRMATORY_COLLECTION", "supersedes_for_current_design": "protocol-v0.8.json", "preserves_history": true, "research_position": "instrumentation-first; T^FIELD is one candidate method, not the premise of the study", "field_types": ["FIELD_TOK", "FIELD_SEM", "FIELD_TRAJ", "FIELD_VIS"], "metric_spec": "/T-GPT/data/metrics-v0.8.1.json", "prompt_set": "/T-GPT/data/pilot-v0.8.1-prompts.json", "annotation_schema": "/T-GPT/data/pilot-v0.8.1-annotation-schema.json", "claims_ledger": "/T-GPT/data/claims-ledger-v0.8.1.json", "freeze_manifest_template": "/T-GPT/data/freeze-manifest-v0.8.1.template.json", "primary_metrics_only": ["ES_LOSS", "REC_GAIN", "IRREV_OBS", "AGG_COVERAGE", "D_sem", "H_sem", "P", "Q_task", "Q_fact", "cost"], "archived_or_exploratory_metrics": ["M_survive", "C_conv", "P_close", "B_pair", "S_first", "U_exp", "T_signature"], "phases": { "development": { "n_per_cell": 24, "purpose": "codebook construction, assignment-rule validation, task-specific Q_task/Q_fact definitions, debugging only", "excluded_from_confirmation": true }, "freeze": { "required_before_confirmation": [ "completed immutable freeze manifest with unique freeze_id", "semantic codebook and cluster IDs", "assignment algorithm and thresholds", "relevance codebook", "primary comparison and endpoint", "task-specific quality rules", "model/checkpoint identifiers", "sampling parameters", "token budget", "exact prompt/probe file SHA-256", "analysis script commit SHA and file SHA-256", "successful deterministic self-test" ], "selftest_expected_output": "T_GPT_V0_8_1_SELFTEST_OK", "changes_after_freeze": "require a new freeze_id and explicit versioned deviation; no silent mutation" }, "confirmation": { "n_per_cell": 64, "purpose": "primary inference; frozen codebook cannot be merged/split/extended" }, "replication": { "n_per_cell_min": 128, "purpose": "independent replication of effects judged important after confirmation; frozen codebook remains unchanged for primary replication" } }, "budget": { "primary_matching_variable": "total input + output tokens", "also_report": ["input_tokens", "output_tokens", "wall_clock_latency", "serial_depth", "parallelizable_calls", "estimated_monetary_cost"], "same_frozen_pool_for_aggregation": true }, "semantic_clustering": { "primary": "frozen bidirectional-entailment semantic identity validated blindly", "sensitivity": ["embedding clustering", "tau_support 0.01", "tau_support 0.03", "tau_support 0.05"], "post_freeze_new_clusters": "UNMAPPED for primary analysis; exploratory extension only" }, "annotation": { "condition_masked": true, "model_masked_where_feasible": true, "T_vocabulary_forbidden_in_codebook": true, "primary_author_as_reference_annotator": false, "minimum_independent_annotators": 2, "agreement": "Krippendorff alpha reported by field", "adjudication": "third blinded adjudicator for disagreements affecting primary endpoint", "confirmation_codebook_mutation_forbidden": true }, "experiments": [ { "id": "E1_STAGE", "priority": 1, "question": "At which post-training transition does observable semantic contraction appear, and what part remains generically recoverable?", "preferred_open_trajectory": ["BASE", "SFT", "DPO", "RLVR"], "template_ablation": true, "primary_endpoint": "adjacent-stage change in ES_LOSS and REC_GAIN", "guardrails": ["Q_task", "Q_fact", "cost"], "minimum_families": 2, "preferred_family": "Tulu/Open-Instruct trajectory where public checkpoint provenance permits stage attribution" }, { "id": "E2_RECOVERY", "priority": 2, "question": "When relevant semantic support disappears in a narrower condition, can generic preregistered probes recover it without naming missing content?", "primary_endpoint": "REC_GAIN", "reference_condition": "preregistered wider condition", "contracted_condition": "candidate narrower condition", "probe_file": "/T-GPT/data/pilot-v0.8.1-prompts.json", "failure_condition": "REC_GAIN does not exceed hostile non-T baselines at matched budget or gains are offset by quality/cost loss" }, { "id": "E3_TOPOLOGY", "priority": 3, "question": "At matched budget and frozen initial branches, which communication/aggregation topology best preserves relevant semantic coverage and provenance?", "generation_pool": "identical frozen initial pool across aggregators", "conditions": ["INDEPENDENT_NO_CONTACT", "IMMEDIATE_DENSE_CONTACT", "TEMPORARY_ISOLATION_DELAYED_CONTACT", "STANDARD_SUMMARY", "MAJORITY_OR_JUDGE", "EXHAUSTIVE_NON_T", "GREEDY_COVERAGE_BASELINE", "T_WEAVE"], "primary_endpoint": "AGG_COVERAGE", "secondary": ["MINORITY_RECALL", "cost"], "hostile_baseline_script": "/T-GPT/tools/coverage_baseline.py" }, { "id": "E4_R_META", "priority": 4, "question": "Does relational topology recur after structural critique more than after equally strong style/length controls?", "arms": ["STRUCTURAL_CRITIQUE", "STYLE_CONTROL", "LENGTH_CONTROL", "NO_CRITIQUE"], "critique_generator": "fixed third-party template/model independent of T^ vocabulary", "primary_endpoint": "blinded structural recurrence rate difference STRUCTURAL_CRITIQUE vs pooled controls", "counter_hypothesis": "CH11_MIRROR_ELICITATION" } ], "hostile_baselines": ["DIRECT", "EXPLICIT_K_DISTINCT", "MINIMAL_MULTI", "DISPERSIVE_SAMPLING", "VERBALIZED_SAMPLING", "MULTI_CLUSTER_GREEDY_COVERAGE", "EXHAUSTIVE_NON_T", "BEST_OF_N_OR_JUDGE", "STANDARD_MULTI_AGENT_DEBATE"], "analysis": { "within_task": "10000-resample non-parametric bootstrap over independent generations", "across_tasks": "task-level bootstrap or preregistered hierarchical model; task is upper-level unit", "confidence": 0.95, "primary_endpoint_per_experiment": 1, "secondary_multiple_comparisons": "Holm within experiment family", "effect_sizes_required": true, "null_results_published": true, "negative_results_published": true }, "auto_null": { "T_layer_removed_if": [ "hostile non-T baseline equals or exceeds T condition within uncertainty at lower or equal cost", "T vocabulary or fixed branch count becomes a detectable stylistic signature without coverage gain", "plurality increases while Q_task/Q_fact loss offsets the preregistered benefit", "benefit disappears under blinded replication" ] }, "reproducible_tools": { "recovery_metrics": "/T-GPT/tools/compute_recovery_metrics.py", "coverage_baseline": "/T-GPT/tools/coverage_baseline.py", "selftest": "/T-GPT/tools/selftest-v0.8.1.py", "tool_readme": "/T-GPT/tools/README-v0.8.1.md", "recovery_example": "/T-GPT/data/example-recovery-v0.8.1.json", "coverage_example": "/T-GPT/data/example-coverage-v0.8.1.json", "legacy_runner": "/T-GPT/tools/run_openai_experiments.py" } }