{
  "id": "MAKE-review-r1-sensor-01",
  "timestamp_utc": "2026-11-09T14:35:00.586928+00:00",
  "status": "Prospective specification for new post-review exploratory analyses on previously opened data; no new confirmation or model promotion",
  "authorization": "Reviewer-driven additional experiments requested by author, delegated within academic-research-suite experiment workflow",
  "cohorts": {
    "hydraulic": {
      "sizes": [
        3,
        4,
        3,
        4
      ],
      "seeds": [
        20751,
        20752,
        20753
      ]
    },
    "temperature": {
      "sizes": [
        10,
        3
      ],
      "seeds": [
        20851,
        20852,
        20853
      ]
    }
  },
  "budgets": [
    64,
    128
  ],
  "prefixes": "Use exact saved fit_ids for every original seed; assert nested prefix, all source, no unseen binding fitting labels; normalize using exact saved prefix mean/std",
  "optimizer": "400 full-batch AdamW updates, lr=.003, weight_decay=.01, float32 neural training, same original initialization seeds and architectures; float64 probability/statistical operations on cuda:3; TF32 disabled",
  "questions": [
    "At fixed pretrained-free forward mean, how do shared shrunken full, diagonal, and identity covariance affect ranking and probabilities?",
    "At fixed joint forward posterior, what changes when replacing it by the product of its exact factor marginals?",
    "With identical factor-one-hot -> SiLU -> sensor-mean decoder and fixed identity Gaussian likelihood, what changes when the decoder training objective is conditional joint cross-entropy instead of observation MSE?",
    "At fixed inverse classifier architecture, does unweighted versus original class-balanced factor cross-entropy account for the gap?",
    "Can a scalar temperature selected using only out-of-fold fitting-prefix predictions address held probability failure?",
    "In hydraulic data, is reversal reproduced under fixed64 normalization/covariance, longer optimization, duplicate64 input, and equal-count additions of new versus already observed bindings?"
  ],
  "main_grid": [
    "MSE forward: full covariance, diagonal, identity; full-covariance posterior projected to product of marginals",
    "CE forward decoder: fixed identity covariance; train conditional CE over full supplied ontology including unobserved bindings without held labels",
    "Inverse MLP: original class-balanced and unweighted factor CE; compare raw posterior and original source per-factor calibration",
    "For full-cov MSE, identity MSE, identity CE, balanced inverse, unweighted inverse: scalar temperature from fixed grid [1,2,4,8,16,32,64] minimizing prefix OOF joint log loss; apply without target tuning; preserve raw outputs"
  ],
  "crossfit": "Four original deterministic folds i mod4, seed+101+fold; one full fit plus four folds for each of four neural objective/architecture choices, 400 steps each. Forward covariance from MSE source OOF residuals with original shrinkage .5/full + .5/diag + .05 I; calibrator OOF forward scores use this same pooled OOF covariance and are not nested. Disclose possible source calibration optimism. No held/source-outside-prefix tuning.",
  "hydraulic_reversal": {
    "normalization_and_inference_covariance": "fixed64 for all controlled refits; use saved64 source OOF covariance; unchanged uniform prior",
    "scenarios": [
      "saved64/400",
      "duplicate64/400 (128 duplicated rows, no new information)",
      "64/800",
      "actual128/400 fixed64 normalization",
      "actual128/800 fixed64 normalization",
      "64 plus q old-binding extra rows/400",
      "64 plus q new-binding extra rows/400"
    ],
    "matched_additions": "Partition later64 rows by joint-binding membership in first64; q=min(partition sizes); preserve prefix order, use first q each. Same initial seed and number of updates. This controls added count, normalization, covariance and initialization; row values and binding identities still differ. All rows inside original128 prefix."
  },
  "outcomes": "Every seed retained. Joint Top1/Top3 with original tolerance1e-12 uniform tie credit, NLL stable log-softmax, multiclass Brier, confidence, ECE10, coverage and accuracy at .9; source OOF and held seen/unseen, temperature each day9-12 and equal-day mean, hydraulic record mean. Standardized sensor-mean MSE and Gaussian observation scores; original source outside128 diagnostic without fitting. No confirmatory p-values or favorable selection.",
  "replay_and_cost": "Save fit_ids, normalization, all full/fold network states, derived mean/covariance, OOF/held log posterior and per-record scores, scalar temperatures, losses at updates0,100,200,400 (and800), raw/source test diagnostics; record all fit/prediction/calibration/serialization times, trainable counts, forward executions, approximate dense arithmetic counts separately from latency, peak GPU bytes and complete process time. No backbone or CPU scientific execution; no offload or DeepSpeed.",
  "checks": [
    "Gaussian CE decoder identity likelihood score matches frozen posterior scores up to additive normalizer",
    "factor-marginal projection normalizes and preserves each factor marginal",
    "duplicate-row MSE gradients agree within float32 tolerance",
    "all original MSE and balanced inverse refit full predictions match saved checkpoints within1e-5 (report residuals)",
    "all reported scalars finite except empty-stratum values explicitly null",
    "all original seeds/prefix counts and original held targets retained"
  ],
  "limitations": "Post-review, opened-data mechanism probes, no independent systems or new confirmation. Objective can be isolated within fixed decoder only; pure modeling direction remains entangled with input/output parameterization. Product projection removes posterior dependence without retraining and does not isolate all factorization assumptions. Temperature calibration cannot establish operational reliability or extrapolate to new platforms.",
  "command": "PYTHONDONTWRITEBYTECODE=1 python articles/make_1/reproduce/review_r1_sensor.py --mode run --device cuda:3 --output articles/make_1/metadata/review_r1/sensor/run_01",
  "hard_timeout_seconds": 1800
}