Detection golden baseline

experiments/detection-golden.json defines a public reconstruction of the first annotated-period membership run. It records the exact recipe configuration, dataset evidence fingerprints, target and preprocessing contracts, sampling policy, and the metrics observed in the historical run. It does not contain recordings, subject identifiers, predictions, model files, or local paths.

The historical run used sleepKIT commit 5c0e5ea8f4693d29972dcbcd1400c025a7b492a0, Python 3.12.5, TensorFlow 2.21.0, Keras 3.15.1, NumPy 2.1.3, h5py 3.13.0, and LiteRT 2.2.0. Its model used the default merged build_model, context 240, five epochs, batch size 32, learning rate 0.001, and seed 0. The final epoch was evaluated once; no early stopping or test-driven selection was used.

This is a reconstruction definition, not a universal runner or a release manifest. The ordinary Python entry point validates the supplied local evidence against the pinned aggregate hashes and then calls run_membership:

Terminal

python experiments/run_detection_golden.py \
  --data /path/to/original-cmidss \
  --events /path/to/train_events.csv \
  --alignment /path/to/source-event-alignment.json \
  --coverage /path/to/candidate-coverage.json \
  --frozen-split /path/to/frozen-split \
  --output /path/to/new-golden-run \
  --cache /path/to/local-feature-cache

The source adapter validates the event, clock, candidate-label, split, and HDF5 evidence before training. The script additionally checks the public definition’s target, context policy, partition sizes, source inventory fingerprint, and all aggregate evidence hashes. Before training it also checks the exact v3 preprocessing SPEC fingerprint and the reviewed preprocessing and model module hashes. The output must be a new directory. Local run output contains the subject-level and prediction evidence needed for review and should remain outside the repository; the script adds golden-definition.json and golden-runner.json so the public definition and runner hash remain attached to that run.

Performance measurements are a separate concern; use the bounded ordinary-Python profiler at python -m sleepkit.recipes.detection.profile with the same local dataset evidence; see the profiling guide. Its results are not part of this golden quality reference.

The target is sleepkit.annotated_period_membership/v1: class 0 is OUTSIDE_SUPPORTED_ANNOTATED_PERIOD, class 1 is INSIDE_ANNOTATED_PERIOD, and -1 is unknown. It represents derived membership in supported annotated nightly periods. It does not assert clinical sleep or wake, confirmed wear, or the competition’s event average precision. The source candidate policy uses complete clock-aligned onset/wakeup pairs and bounded wake periods between eligible consecutive nights; conflicts, unlocatable evidence, and unknown periods remain excluded.

Preprocessing uses TS, ENMO, and ZANGLE at 0.2 Hz. A feature window has 12 samples, advances six samples, and ends at its last source sample. The five features are mean(cos(2*pi*TS/86400)), ENMO mean and population standard deviation, and ZANGLE mean and population standard deviation. Normalization is fitted once from every valid feature frame in the 193 training series, including unknown-target periods and incomplete context tails. Each feature uses the global training mean and sqrt(global population variance + 1e-6). Complete model contexts contain 240 consecutive native feature frames, are nonoverlapping, and are dropped when any feature is invalid or any target is unknown. Removed contexts are never stitched; incomplete tails are dropped.

Training uses all 34,474 eligible native training contexts exactly once per epoch, with no balancing or replacement, shuffled with the declared seed and a bounded buffer of 256. Batches retain the final partial batch. Validation has 8,173 eligible contexts and test has 8,582; both preserve declared subject order and native context order without shuffling. Evaluation covers every eligible test context and 2,059,680 feature frames. The complete context spans 7,230 seconds and uses future features relative to each target, so this is offline, noncausal inference.

The historical test result was 95.7657% accuracy, 0.955611 macro-F1, and 0.127084 cross-entropy. Its confusion matrix, using actual classes as rows and predicted classes as columns, was:

	Outside	Inside
Outside	1,207,331	43,875
Inside	43,338	765,136

These are observed reference values from one recorded run. They are not prospective acceptance thresholds or quality gates. Seeded execution is not a bitwise reproducibility guarantee across environments, and a run from the current checkout is expected to be a reconstruction rather than a byte-identical replay. The first experiment report and annotated target adapter provide the surrounding evidence and interpretation. Exposure of a separate historical SD-2 comparator to this cohort remains unresolved; that question does not qualify this newly trained frozen-split baseline.

The vectorized preparation revision advances only the current preprocessing implementation pin. The original module hashes remain under historical_run, and the original definition remains available at commit 042c1eef5608db3bb842f29c6f12b768663f9026. The preparation equivalence report binds the revised implementation to exact scalar-reference comparisons. Historical quality metrics remain observations from the original five-epoch run.