{
  "schema": "scent-training-contract-v1",
  "status": "Proposed contract. Exposure emulator and linked field model have not been trained.",
  "datasets": [
    {
      "name": "preference_observed",
      "unit": "one observed person-odour response",
      "fields": [
        "dataset_id",
        "person_id",
        "odour_id",
        "pleasantness",
        "calibration_flag"
      ],
      "target": "observed pleasantness",
      "split": "Nested person-held-out validation; additional odour-held-out validation for unseen odour claims",
      "forbidden": "Do not fill missing ratings or reuse resampled visits as independent subjects."
    },
    {
      "name": "exposure_simulated",
      "unit": "one simulation condition, timestamp and zone",
      "fields": [
        "simulation_case_id",
        "door_schedule_id",
        "source_position_m",
        "source_on",
        "time_s",
        "D_m2_s",
        "supply_m3_h",
        "general_extract_m3_h",
        "local_extract_m3_h",
        "boundary_state",
        "zone_id",
        "zone_e",
        "local_removed",
        "other_removed",
        "stored"
      ],
      "target": "Simulated exposure and removal; not physical measurement",
      "split": "Hold out complete design/operation conditions, never random timestamps from the same condition",
      "baseline": "Interpolation of training conditions; direct simulator reference",
      "gate": "Pre-register error thresholds on zone exposure and mass fractions before test evaluation; report zero and near-zero handling",
      "available_seed_cases": 12,
      "missing_for_generalisation": "Current seed cases keep ventilation and geometry fixed; expand operating range before training."
    },
    {
      "name": "choice_future",
      "unit": "consented person-session observations",
      "fields": [
        "person_id",
        "session_id",
        "time",
        "measured_exposure",
        "intensity",
        "comfort",
        "room_choice",
        "dwell",
        "exit"
      ],
      "target": "Observed comfort and choice",
      "status": "Missing",
      "split": "Hold out people and sessions; no bootstrap duplicates across splits"
    }
  ],
  "no_fieldwork_next": [
    "Expand a simulator design of experiments while retaining provenance and full condition groups",
    "Reserve a condition-level test set before model selection",
    "Compare simple emulator to interpolation; report failures as failures",
    "Test public-rating calibration counts separately",
    "Propagate uncertainty through scenario utility, without inventing comfort labels"
  ],
  "current_next_experiment": "ml-next-experiment.json",
  "note": "Experiment 04: residual ML improves overall log error on new general and P10 tests, but Room 03 mean improvement remains uncertain and general tail guard fails. Scaled interpolation stays default. Checkpoint selection fixed; actual one-context residual fit over fixed data.",
  "current_experiment_results": "ml-v4-results.json",
  "live_workbench": "../lab.html"
}
