{
  "protocol": "tab_namer_sealed_final_gate_v1",
  "frozen_at_utc": "2026-08-15T18:26:07Z",
  "repository_commit": "931e57bec0756306651f8033fa606e99ac9aaa98",
  "purpose": "One-shot evaluation of the already-selected B-9500 source-word beam-4 system on the preregistered sealed 1,000 tasks.",
  "selection_complete_before_open": true,
  "no_training_or_parameter_selection_after_open": true,
  "sealed_tasks": {
    "path": "data/derived/campaign/so_clean_holdout_v1/tasks_final.jsonl",
    "rows": 1000,
    "sha256": "4694f1e4b5269b525fd8fb595f45b9ce65d5f831fdd981537da07479831b6874"
  },
  "primary_system": {
    "name": "ours_locked",
    "checkpoint": "output/gsg_scaleup_20260814/keepers/B/checkpoints/step_9500",
    "model_sha256": "0f45a1fef0b549bac512243df0aadad64aedec2a6ade7c17af730247b0b3d52a",
    "config_sha256": "bd146261ee5eaab464002123f5680cc17a0547daee9397fc2d54ec8e8870b8d9",
    "decode": {
      "constraint_mode": "source-words",
      "strip_wrapper": true,
      "max_content_words": 2,
      "max_source_length": 160,
      "max_source_words": 96,
      "max_new_tokens": 8,
      "num_beams": 4,
      "length_penalty": 1.0,
      "generic_title_ban": false
    },
    "glue": "locked Hybrid A with frozen centroid_v2"
  },
  "secondary_product_contract_system": {
    "name": "ours_bounded32",
    "definition": "The same frozen decode and centroid, with the pre-scored whole-word Hybrid-A 32-character bound.",
    "max_title_chars": 32
  },
  "benchmark": {
    "names": ["flan_locked", "flan_bounded32"],
    "checkpoint": "output/student-flan-t5-run5-capacity/best",
    "description": "Campaign title-tuned FLAN-T5-small descendant; not stock FLAN.",
    "model_sha256": "b798d78f3ffcf83510b4e0af55388c75f4232fefbe1a3db17191cfe15b1f902d",
    "config_sha256": "de0d65629d8d29930d8f62cc62e435ecfaddd771392ba8983c9e66eb024c7460",
    "decode": {
      "max_source_length": 96,
      "max_new_tokens": 12,
      "num_beams": 1,
      "do_sample": false
    },
    "glue": "Same frozen centroid_v2 and respectively locked or bounded-32 Hybrid A."
  },
  "centroid": {
    "encoder": "B-9500 input embedding table",
    "idf_corpus": "data/derived/campaign/so_title_gsg_idf200k_v1/train.jsonl",
    "idf_corpus_sha256": "e0f16b6f9162bf65fddb0321637f90f6e0cc70a9bda1e270d8bda57c8a5e02bd",
    "idf_limit": 200000,
    "k": 3,
    "mmr_lambda": 0.7,
    "min_df": 0,
    "idf_floor": 0.0,
    "idf_cap": 0.0,
    "drop_pure_digits": false,
    "drop_long_hashes": false,
    "require_mapped_words": false
  },
  "shared_assets": {
    "tokenizer_json_sha256": "a5a4877b870af92adaa0ec9416659e288915a916c90854623ec6b41cc70ed02f",
    "tokenizer_config_sha256": "a31d846bdb8ccdd09149192ab55b81b88ab2fc240ddb2a0bbbd5f79f0e205100",
    "vocab_map_sha256": "87ca97e2b3e95322143caa24f6a841321aa15dca8ea40c6f75648d6cc3337f80"
  },
  "implementation_sha256": {
    "scripts/generate_constrained_titles.py": "0e3f08965a17db494229bff00e2ab9d54aaf85e8b10d17ea6aa20add83716384",
    "scripts/generate_title_candidates.py": "314d8a4c67e14b162a9fd665533db4b8431dca36beb0eb566aa36ccbceb2225f",
    "scripts/centroid_headline_v2.py": "d9ae2c8ba88457abf9e8f131d2c4dbed62d8c2b5e59f000ba513196438deef5b",
    "scripts/glue_namer_centroid.py": "56f4f13a8471f7f8fc855bc5f7e2abb49448f5733524f0371a329638da2a0880",
    "scripts/judge_title_candidates_gemini.py": "44fa1902af81792af0686952ca692bebc954d2c678cbeeeb2b3c047707a370e2",
    "scripts/build_blind_title_review.py": "00d588c4d1ebd95878ba1bc3522da2a6ad64f03d522f03c736e0060e4b02f9a9",
    "scripts/summarize_blind_title_review.py": "75de36fee513c1c259145cc7cb7edda9a6f3c6a349a1aeb137c1fcab0471d232",
    "data/vocab_map.py": "4a4af6c5d35aec31a8f8f5870e402ea33f02f68e452419ffa6c83c1b65591ef9",
    "data/title_rules.py": "0c22611b03fb60263513f9379d689dc08e25d384a7bbf9fffcb9bcfa42a938c1",
    "train/hybrid_glue.py": "ac3a41277fd6eed04f2aa2e565edd5d2b530ae806570636bed45fb1173b5e8c3",
    "scripts/centroid_headline.py": "d7746f60f46f1b42393d28b3ccb2fa64e4a45ec2402ffd059d006ad24b94fc78",
    "scripts/build_extractive_titles.py": "e96211a9ec84dbf0dad548a45e59e6c675fd222a4ebf3a07b67f70ee93feb84b"
  },
  "preflight": {
    "clean_dev_tasks_sha256": "3261f1c459275e934e7fde2ffd100f378d48286f8e03d61757726a15803ebb8f",
    "ours_raw_replay_sha256": "f736d9c1e232d99fc56d21bc208004c998c576e0aea085d2fd7742619dd90477",
    "ours_locked_replay_sha256": "5f95a123491fd1c8f9216ed149db57ac62f9bd8638ec0b564e2aa9b07cd471ca",
    "flan_raw_replay_sha256": "9223a8e580e204601e350177e4d2500cbbc73f31993479e96228068177b2d326",
    "flan_locked_replay_sha256": "b6d94bb4d5c1a913dd825a80673d9158e6b622f5d38b4bb6a8c315ce7fb529b9",
    "status": "All generation and glue predictions replayed exactly before the sealed run."
  },
  "gemini_packet": {
    "single_packet": true,
    "candidates": ["ours_locked", "flan_locked", "ours_bounded32", "flan_bounded32"],
    "model": "gemini-3.5-flash-lite",
    "temperature": 0,
    "seed": "tab-title-gemini-judge-v1",
    "batch_size": 20,
    "candidate_order": "Per-row deterministic SHA-256 shuffle.",
    "identity_blinded": true,
    "reference_labels_shown": false,
    "primary_metric": "Mean contextual tab-title usefulness on the 1-10 scale, two decimals.",
    "paired_primary": "ours_locked minus flan_locked",
    "paired_secondary": "ours_bounded32 minus flan_bounded32"
  },
  "independent_audit": {
    "sample_size": 40,
    "seed": "tab-title-final-gate-codex-audit-v1",
    "candidates": ["ours_locked", "flan_locked", "ours_bounded32", "flan_bounded32"],
    "identity_blinded": true,
    "reference_free": true,
    "score_scale": "0-4; useful means 3 or 4",
    "judge": "Codex, without access to mapping_private.jsonl until judgments are complete"
  },
  "preregistered_gate": {
    "primary_mean_at_least": 5.5,
    "primary_audit_useful_rate_at_least": 0.6,
    "primary_audit_wins_plus_ties_vs_flan_at_least": 0.3,
    "primary_empty_rate_at_most": 0.01,
    "primary_top_title_mass_at_most": 0.1,
    "stop_recipe_if_primary_mean_below": 4.5
  },
  "descriptive_analysis": {
    "paired_mean_delta_bootstrap": "Deterministic percentile 95% CI, 100000 paired row resamples, seed tab-namer-final-gate-bootstrap-v1.",
    "rate_intervals": "Wilson 95% intervals.",
    "selection_or_gate_use": false
  },
  "runtime": {
    "python": "3.11.13",
    "torch": "2.13.0",
    "transformers": "5.15.0",
    "tokenizers": "0.22.2",
    "generation_device": "Apple MPS"
  }
}