{
  "experiment_start_utc": "2026-09-10T10:57:23Z",
  "hard_deadline_utc": "2026-09-10T14:57:23Z",
  "student_initialization": "All weights originate from random initialization during this experiment. Copy-head parameters were added later; no pretrained student weights were used.",
  "teacher_role": "Language templates, direct paraphrases, and round-trip verification; reference SQL semantics constructed programmatically.",
  "curriculum": [
    {
      "data": "v1",
      "examples": 85777,
      "tokens": 22739579,
      "purpose": "Initial semantic and language learning"
    },
    {
      "data": "v2",
      "examples": 152667,
      "tokens": 43182146,
      "purpose": "Random character identifiers and tool names"
    },
    {
      "data": "v3",
      "examples": 218022,
      "tokens": 62519864,
      "purpose": "Training-vocabulary chunks and varied MCP naming styles"
    },
    {
      "data": "v4",
      "examples": 283377,
      "tokens": 81750566,
      "purpose": "Identifiers without fixed prefixes; copies restricted to context/question",
      "sample_weights": {
        "latest_65355_variants": 3,
        "other_rows": 1
      }
    },
    {
      "data": "v5",
      "examples": 302893,
      "tokens": 87335945,
      "purpose": "Exact copying of project IDs and schema discovery table names",
      "sample_weights": {
        "scope_and_discovery_variants": 4,
        "unprefixed_identifier_variants": 3,
        "other_rows": 1
      },
      "validation_selection": "All 1,200 validation records, token-weighted loss"
    },
    {
      "data": "v6",
      "examples": 320115,
      "purpose": "New schema families with related table/parent/project identifiers; authored comparison contrasts; conservative training-template audit filtering",
      "new_scenarios_before_filter": 6000,
      "new_rows_before_filter": 24000,
      "excluded_rows": 6778,
      "validation_selection": "All 1,200 validation records, token-weighted response loss",
      "tokens": 92098292,
      "response_tokens": 12137844
    },
    {
      "data": "v7",
      "purpose": "Variable schema layouts, random qualified tool names, authored natural city requests and Hindi lexical mappings",
      "new_layout_variants": 8826,
      "new_runtime_name_variants": 13635,
      "new_city_language_rows": 4800,
      "new_city_scenario_families": 1200,
      "new_variant_sample_weight": 6,
      "development_observation": "An informal default-schema demo failed on a shorter schema and natural language. This stage addresses that development failure; the demo is not a held-out benchmark.",
      "examples": 347376,
      "tokens": 100224712,
      "response_tokens": 13163249
    }
  ],
  "implementation_commits": [
    "f24d869",
    "c1ce778",
    "c9e8d77",
    "f6f0b7b",
    "780db6a",
    "4875531",
    "f713a08",
    "69e19ff",
    "1f798c7",
    "47267b7",
    "5d7a622",
    "3ebf37a",
    "eecb176",
    "1c50583",
    "364753b",
    "a8f336f",
    "1049d88",
    "9729c63",
    "369909b"
  ],
  "architecture_reference": "https://aclanthology.org/P17-1099/",
  "research_claim": "Established decoder components and a learned pointer-generator adaptation. No claim of a new research architecture or controlled ablation.",
  "template_audit": {
    "teacher": "Qwen/Qwen3.8-27B-FP8",
    "mtp": 3,
    "jobs": 330,
    "template_instances": 1313,
    "accepted": 1276,
    "flagged": 37,
    "training_rows_excluded": 6778,
    "scope": "Training SQL templates only; held-out validation/test/manual phrases excluded",
    "caveat": "Earlier stages used pre-filter corpus; final continuation uses filtered data"
  },
  "manual_split": "Additional hand-written phrasing templates rendered on 40 test scenario families; no shared train/validation scenarios. Manual and test are not independent schema-family samples.",
  "development_split": "192 schema/tool-name perturbations of 140 validation families; no training, test or manual families. Used for development, not an independent test.",
  "selection": {
    "final_rule": "Among frozen evaluated candidates, maximize the mean of full validation task success and development task success; use validation response loss to select additional candidate snapshots. Final test and manual results are read only after weights are selected.",
    "rationale": "Standard validation alone missed the schema-layout failure found in the development demo."
  },
  "refinement": {
    "data": "v7, unchanged",
    "start_utc": "2026-09-10T13:40:00Z",
    "planned_minutes": 18,
    "batch": 128,
    "learning_rate": 0.0001,
    "prompt_weight": 0,
    "response_weight": 1,
    "action_auxiliary_weight": 0.05,
    "rationale": "Focus late optimization on output actions after earlier language learning; reduce unpredictable-input LM gradients. Selection can retain an earlier checkpoint if refinement does not improve development scores."
  },
  "selected_checkpoint": {
    "step": 25715,
    "sha256": "9489b5cde69c11f64b3e7031182e8248aa83790f1ad4aeefbf99a4e5e152e692",
    "frozen_at_utc": "2026-09-10T14:00:36.513800+00:00",
    "test_and_manual_excluded_from_selection": true
  },
  "selected_checkpoint_counters": {
    "step": 25715,
    "processed_tokens": 595067301,
    "response_tokens": 78367410,
    "training_seconds": 5092.33953666687,
    "random_initialization": true
  },
  "completed_trainer_counters": {
    "step": 32642,
    "processed_tokens": 851198143,
    "response_tokens": 113051294,
    "training_seconds": 7099.562133073807,
    "note": "Retained-lineage trainer counters include validation/checkpoint time, exclude discarded work/model loading. Last checkpoint was not selected."
  },
  "release_archive_limitation": "SSH briefly refused connections, then recovered. Runtime configs, training logs, candidate/selection records, GPU predictions and native evaluation reports were retrieved. The optional 1.67 GB final optimizer-state transfer was stopped to avoid extending rental cost; selected inference weights are fully secured."
}