Event data architecture¶
Immutable CAN recordings and reviewed manifests are the source of truth.
graphids.data.recordings validates explicit source roles, hashes,
clock units, payload provenance and supervision. Reviewed clock repair mappings
remain external artifacts. Ingestion preserves physical order and rejects
decreasing clocks; it never silently sorts or repairs a source.
can_event is the only active representation. can_event_dm writes lossless
event-v1-<recording-digest> caches with unchanged content identities: FP64
physical times, all 29 ID bits, uint8 payload bytes, lengths and unknown labels.
Only selected source queries are opened. Reserved target recordings and reference
roles never become training queries.
EventStreamWindows schedules chronological chunks from independent recordings.
It groups equal lengths, carries separate row positions, processes short tails,
and replaces finished streams without padding, replay or dropped scored rows.
EventTrainingWindow declares complete classification and normal-payload target
totals for one optimizer update. The model detaches history between chunks.
graphids.evaluation.EventScores writes one bounded frame-score shard per
chronological chunk and verifies exact recording coverage before publishing the
index. Epoch AP consumes scalar columns via a disk mapping; node histories and
score blocks are not retained by the model. This artifact component has no torch
or Lightning dependency.
See model and training contract and the retired snapshot archive. The experiment CLI, Ray, Lightning and Slurm infrastructure retain their existing roles.