| { |
| "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 |
| } |
|
|
