Run records and model cards

results/
  runs/<id>.json        one record per formal run, named after its model card
  methods/<id>.json     one model card per method
  arena/<entry>.json    one disclosure card per arena entry

Run records

python -m nexttransformer.export_run writes a record from a finished run directory and its audit; python -m nexttransformer.export_run --check validates every record in CI. The exporter refuses a run unless all 189 validation and six full CORE-22 events are present, the per-task CORE scores reproduce the aggregate, the cumulative audit is matched, every segment was formal, and the run used the pinned tokenizer and the frozen protocol.

Field	Meaning
id, methodId, cohort, name	Run id, registered method (SSSL keeps dense), accounting group for the checks, display name
status, step, complete	finished or partial; last update; whether the full protocol was met and accepted
val, core	{step, value} points: validation BPB at every scheduled step; CORE on its 0–1 scale
taskEvaluations	Per CORE event: raw accuracy and centered score for each of the 22 tasks
finalBpb	The independent final validation BPB, reported separately from the curve
trainingCompute	Cumulative training FLOPs, average GFLOPs per input token, and the accounting used
proof	Protocol, model, tokenizer, data and source fingerprints of the run
source, url	Provenance of the record and its evidence in the repository
comparisonNote	A provenance note shown with the run, when needed
baselineAcceptance	Only on the accepted dense_full run

cohort names the accounting the checks apply: reference-v1 (the operator-level ledger), native-v3 (the historical matrix-FLOPs estimate) or frozen-replay-20261019 (two runs with their original replay accounting). It does not select a separate leaderboard. The file name is the model card id; --file NAME distinguishes a second run of a method.

Model cards

methods/<id>.json holds the name, family, tagline, description, paper and code links, mechanism, limitations, parameter and block counts, configured compute, and the architecture diagram the observatory draws. Its id is the methodId of run records. "status": "pending" marks a registered method without a published run; the observatory lists it with a “Results coming soon” badge and never ranks it.