Predictive representations for compact Bayesian decision making
Reproduction assets for Predictive Representations for Compact Bayesian Decision Making: Controlled Compression and Compositional Sensor Diagnosis.
Includes all 132 sensor-model checkpoints (11 methods, two label budgets, three seeds, two datasets), processed features and split indices, generated probability laws and samples, complete reference scores, fixed protocols and executable code. Checkpoints use paired JSON and Safetensors files. No pretrained language model or executable pickle checkpoint is required.
Run
Python 3.11+ and a CUDA-enabled PyTorch installation are required. The recorded environment is PyTorch 2.6.0/CUDA 12.4 on an NVIDIA H200; timings depend on hardware.
python -m pip install -r requirements.txt
python reproduce.py --mode check --output checksum_report.json
python reproduce.py --mode all --device cuda:0 --output replay_report.json
python analyze.py --device cuda:0 --output statistics.json
python paper/build_figures.py
python postconfirmation.py --device cuda:0 --output postconfirmation_replay
Replay checks all 132 checkpoints, the controlled primary comparison and the four-day sign-flip result. Analysis recomputes paired uncertainty from the supplied reference scores. Figure generation writes all article figures and tables under paper/.
Optional independent refitting, source reconstruction and sensitivity checks:
python reproduce.py --mode refit --device cuda:0 --output refit_report.json
python rebuild_sources.py --device cuda:0 --cache raw_downloads --output source_rebuild.json
python rebuild_external.py --device cuda:0 --output external_reproduction
python rebuild_controlled.py --device cuda:0 --output controlled_reproduction
python paper/sensitivity_analysis.py --device cuda:0 --output sensitivity.json
Raw-source reconstruction downloads the official archives and verifies their digests. External and full controlled reruns can take substantially longer than checkpoint replay. Use a new output directory for each rerun. Original arrays and numerical modules are unchanged.
Contents and scope
- models/: every fitted sensor comparison, including adverse results.
- data/: processed sensor inputs, labels, splits, generated laws and sample streams.
- results/: reference predictions, complete external metrics and controlled frontiers.
- protocols/ and provenance/: fixed recipes, source definitions, eligibility failures and measured compute.
- unionmind/: numerical implementations; use the standalone commands above.
- SHA256SUMS.json: integrity hashes for the distributed files.
The original six-control compression comparison gains 0.1013 utility at 1024 records. Against the stronger action-value encoder, the primary-mixture gaps are 0.00194, 0.01935, 0.01397 and 0.00086 at 64, 256, 1024 and 4096 records; this added comparison is descriptive. The external compression improvement criterion fails. The four-day temporal comparison is secondary because one source day fails its clock check; exact one-sided day-level p=0.0625. The 128-label hydraulic reversal, calibration failure and lack of complete-cost superiority are retained. These artifacts do not establish operational root-cause reliability.
Additional analysis on opened data
An action-value compression control, a complete 1,212-episode partition and normalization checks for all 132 checkpoints, and crossed frozen hydraulic means/covariances are in results/postconfirmation/. The single added encoder was specified after the original confirmation was opened. Its specification and code hash are in protocols/postconfirmation_analysis_01.json. Original protocols, promotion decisions and checkpoints are unchanged. The command above recomputes the complete analysis without refitting. Measured timings are hardware-dependent.
Data and license
Code is MIT; original models, generated data and research outputs are CC BY 4.0. Hydraulic features derive from Helwig, Pignanelli and Schuetze (UCI DOI 10.24432/C5CW21); temperature-modulated sensor features derive from Burgues (10.24432/C5S302). Gas-drift recordings (Vergara, 10.24432/C5RP6W) are retrieved separately and are not redistributed; preserve the source's legacy research-only notice. Full attribution and terms are in LICENSE.md.
Reviewer-revision experiments
revision_r1/ adds exploratory mechanism tests on the already opened data:
full controlled action-value/loss comparisons, external sample-support and
fixed-policy transfer comparisons, and 318 sensor fits with all seeds and
budgets retained. Original results, protocols and promotion decisions are unchanged.
The 27 sensor state bundles use only Safetensors and JSON. Specifications,
all comparative scores, cost records and numerical tie-sensitivity diagnostics
are included; these are additional analyses, not new confirmation.
Replay every added sensor state without fitting:
python revision_r1/sensor/verify_all.py --release . --base revision_r1/sensor/run_01 --extension revision_r1/sensor/objective_extension_01 --output revision_sensor_replay.json --device cuda:0
Recompute the controlled interventions in a new directory:
python revision_r1/controlled/run.py --prepare --release . --output new_controlled --device cuda:0
python revision_r1/controlled/run.py --release . --output new_controlled --device cuda:0
The external analysis requires separately acquired gas-drift recordings. Retain the provider's research-use notice; no transformed gas records are redistributed.
python rebuild_external.py --output new_external_source --device cuda:0
python revision_r1/external/run.py specify --release . --input new_external_source --output new_external_analysis --device cuda:0
python revision_r1/external/run.py run --release . --input new_external_source --output new_external_analysis --device cuda:0
Sensor refitting is optional through revision_r1/sensor/review_r1_sensor.py
and objective_extension.py, each accepting --release ., --output and
--device. Use a new output directory. Rank gains do not establish calibrated
acceptance, and sparse fitted-law ties can affect ancillary realized utilities;
the unchanged primary mixture frontiers and all disclosed discrepancies remain available.
Additional scientific validation
validation_20261110/ contains the locked new-sensor replication, strict nested
source calibration and optimized Gaussian inference comparison. The new sensor
test passes support checks but fails the preselected neural improvement criterion.
All methods, budgets and negative results are retained. Original artifacts remain
unchanged. Reproduction commands and source terms
include the separately acquired Air Quality archive; no restricted raw or
processed recordings are redistributed. Portable reruns reproduce an existing
study and do not constitute new prospective confirmation.
Stronger-baseline and sample-size analysis
editor_revision_20261110/ supplies the complete fixed-estimator frontier,
authored-family contributions, utility/query-charge decomposition and true-model
boundary calculations. These are descriptive summaries of frozen outputs, not
new fitting or confirmation. Original model files and scores are unchanged.
python editor_revision_20261110/summarize.py --root . --output replay_editor --plot
The numeric summary uses the Python standard library; optional plotting uses
matplotlib. The source manifest identifies archived utilities separately from
later diagnostic replays. Every estimator, sample budget and mechanism family
is retained. Generated controlled.pdf and plot_data.csv reproduce the updated
controlled figure; no raw third-party recordings are needed.