Blame for protocols/day_transfer_protocol_01.json

{
  "id": "day_transfer_01",
  "objective": "Independent, separately registered full-hypothesis cross-day retest of unchanged compositional world models. The preceding strict-clock eligibility failure is final and is not amended or replaced.",
  "source": {
    "uci": 487,
    "doi": "10.24432/C5S302",
    "author": "Javier Burgues, Institute of Bioengineering of Catalonia",
    "url": "https://archive.ics.uci.edu/static/public/487/gas+sensor+array+temperature+modulation.zip",
    "license": "CC BY 4.0",
    "access": "Only official metadata and README accessed before lock; archive and numerical values unopened."
  },
  "source_definition": "13 timestamp-named files in order. First nine days fitting pool; last four days held. Preserve every day; any missing or invalid day fails complete cohort. Native nondecreasing seconds and source row order allowed; repeat timestamps never imply extra independent observations. One 900-second provider protocol block is one episode; require >=850s coverage. At least90 episodes per day,128 fitting records,and8 unseen-binding cases per held day.",
  "features": "Four consecutive 225-second means for each of14 resistance channels (columns6:20),56 values. No CO,humidity,temperature,flow or heater reference enters features. Preserve source MOhm. No feature selection.",
  "ontology": "10 author CO levels uniformly0..20ppm, identified by round(9*median_last_quarter_CO/20),clamped0..9; three measured RH bands <=35,(35,55],>55 percent using last-quarter reference mean. Full30 states. Reference labels are supervision only.",
  "split": "Hold first6 of30 ontology bindings by SHA256(20821:factor0,factor1). Entire binding excluded from first9days fitting pool. Report all held observations, seen/unseen strata, each day and unsupported factor levels; no after-opening selection.",
  "sample_budgets": [
    64,
    128
  ],
  "seeds": [
    20851,
    20852,
    20853
  ],
  "sampling": "Same random fitting prefix across all11 models and both budgets; no class balancing or target tuning.",
  "models": [
    "additive_full",
    "additive_diag",
    "factorized_neural",
    "joint_neural",
    "factorized_nb",
    "factorized_lda",
    "factorized_kernel",
    "factorized_ridge",
    "joint_gaussian",
    "factorized_logistic",
    "factorized_classifier"
  ],
  "candidates": [
    "additive_full",
    "factorized_neural"
  ],
  "algorithms": "Additive: all factor one-hots plus intercept, ridge 8, intercept ridge .01, analytic leave-one-out residuals. Factorized neural: full factor one-hots -> SiLU hidden -> observation means. Joint neural: one-hot complete binding -> matched SiLU hidden -> observation means, unknown bindings use uninformative source mean. Hidden widths give <=20000 counted trainable parameters and near-matched factorized/nonfactorized budgets. Both neural decoders fit 400 AdamW steps, lr .003, weight_decay .01; four source folds estimate held residuals plus one full fit, all charged. Source residual covariance is .5 full + .5 diagonal with .05 identity floor; diagonal ablation uses only the diagonal. Priors uniform over the supplied full hypothesis ontology, including unseen bindings.",
  "controls": "Independent-factor Gaussian NB and shared-covariance Gaussian, independent-factor whitened kernel with bandwidth .7 and .5 Dirichlet smoothing, independent-factor ridge classifier with ridge 8, joint-binding Gaussian with shrink-to-source fallback. Use all observed coordinates and the same fitting labels. All factorized controls can predict unseen bindings; do not make the comparison only against memorization. Also include class-balanced multinomial logistic regression and a matched <=20000-parameter SiLU direct classifier, both predicting independent factor probabilities, with the same 400-step optimizer and four out-of-fitting-fold source predictions plus full fit. Their four-point source temperature grid (.5,1,2,4) is chosen only on source cross-fit log loss; every fit is charged. All source- and target-state priors remain uniform.",
  "normalization": "Feature mean and standard deviation from the actual 64/128 fitting records only. No full-source label-conditional quantiles. Standard deviation floor 1e-6. All models receive identical normalized numerical evidence.",
  "uncertainty": "20000 paired episode draws independently within each of4held days, equal-day weights,seed20891,family40. Additionally paired day-cluster bootstrap; require both intervals positive. Three seed results averaged as independent fits, never a prediction ensemble. Only4helddays: no population-of-systems generalization.",
  "success": "Primary64examples: >=.10 unseen-binding jointTop3 gain over all7strong baselines,positive simultaneous95% lower bound in both bootstrap analyses,positive point gain on each held day. Overall superiority additionally requires positive intervals against all10alternatives. Neural-specific claim requires same against additive reference.128-label result cannot rescue64primary. All13days and complete cost ledger required.",
  "acquisition": {
    "policies": [
      "value_forced",
      "value_stop",
      "variance_forced",
      "variance_stop",
      "static_forced",
      "random_forced"
    ],
    "cohort": "First8unseenbinding held episodes per held day under SHA256(20901:recordID); all primary records remain.64examples,firstseed,2fixed world models.2/4/8queries,unitcost.01,unchanged exact one-step value operations. Secondary only. No acquisition advantage without>=.10gain over allcheap controls and paired simultaneous family120 bounds; actual processing cost separately reported."
  },
  "compute": "CUDA float64 statistics, probability operations, exact interval decisions and bootstrap; float32 auxiliary neural fitting. CPU archive text parsing, provenance, hashing and report aggregation only. Count data preparation, normalization, every full/cross-fit adaptation, full-hypothesis mean materialization, covariance construction, predictions, query decisions, peak GPU memory, trainable/checkpoint/derived-state bytes. No CPU offload, DeepSpeed, Qwen executions or model collaboration.",
  "stopping": "No new model,capacity,threshold,feature,sample,day or cohort revision after numeric access. Code and inherited modules hashed before source download. Preserve failed original source-clock study and every preceding result. No human-review dependency,no Qwen/backbone execution,no article-only output."
}