DATASETS.md

Simulation inputs

This release separates historical originals from reconstructions. All files are supplied, so material reconstruction and checkpoint evaluation do not download hidden inputs.

Dataset	Episodes	Status	Historical container exact
data/paper/test.npz	480	Archived original	Yes
data/paper/train.npz	2400	Archived original	Yes
data/rebuttal/dynamic_frequency_sweep/test_freq_0020mhz.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/dynamic_frequency_sweep/test_freq_0040mhz.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/dynamic_frequency_sweep/test_freq_0160mhz.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/dynamic_frequency_sweep/test_freq_0320mhz.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/dynamic_frequency_sweep/test_freq_0640mhz.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/test_dynamic_slow.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/test_flow_switch_100.npz	480	Archived original	Yes
data/rebuttal/test_magnitude_100.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/test_magnitude_125.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/test_magnitude_150.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/test_magnitude_200.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/test_with_params.npz	480	Exact after metadata restoration	Yes
data/rebuttal/train_with_params.npz	2400	Exact after metadata restoration	Yes
data/rebuttal/validation_dynamic_slow.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/validation_id.npz	480	Recorded-protocol reconstruction	No
data/rebuttal/validation_magnitude_150.npz	480	Recorded-protocol reconstruction	No

Archived originals and the two exact metadata restorations match their historical containers. The thirteen reconstructed inputs follow the fixed generation commands, including paired episode seeds, hidden-flow instances, boat parameters and initial states across stress/frequency conditions. Existing-input design checks pass and are recorded in data/manifest.json.

The original containers of those thirteen derived splits were not recovered. Their equality to the historical numerical trajectories is therefore not asserted. Published tables and figures are rebuilt from the preserved historical per-seed and per-episode results, never from substituted reconstructed-input evaluations. New evaluations record the provenance of the inputs actually used.

The ten base derived splits match the historical recorded shapes, metadata fields and family/trajectory counts. All numerical arrays are finite. The release also checks train/test/validation episode separation, hidden-flow-pool separation, paired stress/frequency conditions, exact labeled/original trajectory identity and the locked switching transitions. These are input checks, not a replacement experiment campaign.

Training contains 2,400 episodes, 720,000 transitions and 722,400 simulator states; the original test contains 480 episodes, 144,000 transitions and 144,480 states. NPZ files hold simulator trajectories and commands. RGB observations are rendered online, without visualization overlays.

Use sh run.sh generate-data --output runs/generated_inputs to create a separate set from the recorded commands. Cross-platform floating-point and serialization equality is not assumed. The existing license state is retained; this update adds no license grant.