{
  "schema": "industrial.operating.challenge.v2",
  "name": "Industrial Operating Evals v1",
  "status": "public development; no independently verified leaderboard",
  "control": {
    "name": "Delayed multivariable control under operating limits",
    "source": "Johansson equation-level adaptation, DOI 10.1109/87.845876",
    "parent_protocol_sha256": "a6639df62fb251e248ab830470ec40ebca214b806ccf0fe1420da8032427b3b5",
    "regimes": [
      "minimum_phase",
      "nonminimum_phase"
    ],
    "core_conditions": [
      "nominal",
      "flow_step",
      "pump_loss",
      "split_change",
      "quality_gap",
      "hidden_initial"
    ],
    "ambiguity_conditions": [
      "bias_plus",
      "bias_minus"
    ],
    "development_seeds": [
      93288101,
      93288102
    ],
    "episode_steps": 480,
    "sample_seconds": 5.0,
    "warmup_steps": 16,
    "admission_consecutive_usable_records": 16,
    "reference_preview": false,
    "measurement_delay_steps": 1,
    "voltage_bounds_V": [
      0.5,
      5.0
    ],
    "slew_V_per_sample": 0.15,
    "constraint_tolerance_V": 1e-06,
    "sampled_level_envelope_cm": [
      0.2,
      20.0
    ],
    "research_bar": {
      "maximum_group_mean_episode_true_RMSE_cm": 0.5,
      "sampled_envelope_violations": 0,
      "invalid_commands": 0,
      "deadline_misses": 0,
      "quality_hold_failures": 0,
      "participant_errors": 0
    },
    "bar_origin": "Organizer-selected research challenge using known historical reference results; not a field requirement or proof that every group is feasible. Frozen before these new development outcomes. Feasibility beyond present references remains open. V1 post-run audit found two exact steady-infeasible requested setpoints among 120 checks. This separate new-cohort v2 generator rejects infeasible targets before outcome access; no operating bar or reference setting is loosened.",
    "process_contract": {
      "transport": "one JSON request / one JSON response per line, persistent subprocess per episode",
      "initialize_deadline_seconds": 20.0,
      "decision_deadline_seconds": 0.5,
      "maximum_response_bytes": 4096,
      "response": {
        "command_V": [
          "finite number",
          "finite number"
        ]
      },
      "allowed_reset_information": [
        "nominal regime",
        "public nominal equilibrium",
        "operating voltage",
        "16 actual warmup records",
        "sample interval",
        "bounds and slew"
      ],
      "allowed_decision_information": [
        "decision index",
        "delayed measurement or null",
        "source sequence",
        "quality",
        "actual previous command",
        "current reference",
        "admission gate"
      ],
      "not_delivered": [
        "seed",
        "condition",
        "true current/upper state",
        "actual changed splits/gains/flows/bias",
        "future target",
        "future noise"
      ],
      "invalid_or_late_response": "count failure; apply previous voltage, never clip an invalid proposed command into eligibility",
      "isolation": "cooperative subprocess contract, not an OS security sandbox; private reserve scoring requires separately verified isolation with no seed/truth mounts"
    },
    "references": [
      "hold",
      "selected_pi",
      "selected_nominal_mpc"
    ],
    "selection": "Reuse original validation-selected settings; no new fit, search or retuning",
    "ranking": "Only candidates passing every core gate may be ordered by worst group mean episode true RMSE, then macro true RMSE, then total command variation. No pooled compensation for a failed group. Failed arms are unranked.",
    "ambiguity_interpretation": "Report true and observed tracking separately. Bias cases do not enter the core passing bar and cannot establish calibration or physical bias-vs-drift diagnosis.",
    "timing_interpretation": "Measured included-host process deadlines and response latency; not a PLC deadline, WCET, target-device result or continuous-time safety proof.",
    "reserve": "Separate random synthetic seed set with salted SHA256 commitment; not executed, not physical independence, no public hidden-scoring service or valid private rank until isolation and independent submissions are verified",
    "target_generation": {
      "name": "common steady-feasible targets",
      "candidate_model_corners": [
        "nominal",
        "nominal_plus_frozen_flow",
        "frozen_pump_loss",
        "frozen_split_change"
      ],
      "reference_offsets_cm": [
        0.75,
        1.5
      ],
      "maximum_rejection_candidates_per_reference": 2000,
      "criterion": "Every target is exactly steady-reachable under all four declared dynamics corners, voltage and all four sampled-level bounds. This does not certify dynamic RMSE, safety or stability.",
      "past_or_future_fault_identity_used": false
    }
  },
  "plume": {
    "protocol": "industrial-plume-operating-budget-v0",
    "dataset_revision": "ed2b750d2fb76c6b9dac9d4e106e6c16390ad571",
    "n_tiles": 342,
    "n_flights": 44,
    "n_dates": 3,
    "primary_bar": {
      "every_date_negative_indication_max": 0.05,
      "weak_localized_recall_at_IoU_0_1_min": 0.8,
      "weak_mean_IoU_min": 0.25
    },
    "status": "Existing retrospective development challenge; released model test-loss monitoring, consumed public labels, later-to-earlier calibration and annotation incorporation bias disclosed",
    "chemical_no_leak_truth": false,
    "submission": "Complete dense finite score grids, original pinned label/source checks and the original participant grader; this release does not claim to have closed its actual participant GDAL execution gap"
  },
  "well": {
    "name": "Causal production-event alarms with explicit normal exposure",
    "model_revision": "0b59a3aee687c3bec7553d4c5c8860a598d7062a",
    "dataset_revision": "342dfb94f4c433de23ffdb770f4ce89ddc0143af",
    "decisions": 80534,
    "research_bar": {
      "false_alarms_per_eligible_normal_day_max": 1.0,
      "normal_alarm_occupation_max": 0.01,
      "known_onset_coverage_within_300s_min": 0.8
    },
    "bar_origin": "The original alarm budgets plus an explicit organizer onset coverage target; new target is retrospective and not a field requirement",
    "eligibility": "Report common/full input eligibility, unknown labels, raw/grid/eligible seconds and censored observable onsets; carried-in alarms are not pre-onset forecasts",
    "status": "Previously consumed initial confirmation, now public development evidence; no new model inference or reserved-group access in this suite",
    "submission": "Original causal score grid and fixed alarm state; this release exports the original executable alarm scorer and a completeness guard, not a new prospective predictive study"
  },
  "compute": {
    "paid": false,
    "gpu": false,
    "tpu": false,
    "controller_CPU_wall_limit_seconds": 1424,
    "new_training": false,
    "models_retuned": false,
    "prior_two_harness_seconds": 374.79147018799995,
    "total_three_harness_cap_seconds": 1800
  },
  "scientific_claims": {
    "independent_community_attempts_verified": 0,
    "official_rank_verified": false,
    "field_qualification": false,
    "difficulty": "Measured failure of named organizer references only; no claim experts cannot ace it",
    "private_well_reserved_groups_opened": false
  }
}