publish-paper

Publish the paper

Needs a main.tex and results with their provenance manifest in a clone. Two stages with two checkpoints: local green (your own check) and CI green (the check outside your hands — only its artifact is the paper of record). A red CI gate loops back to the local stage.

Local: realize the record and write the paper

1. Realize the record from the run outputs. The declarations live in .research/record.json (created at preregistration); update_record reads them and .research/results/ itself — numbers cannot be passed in — appends to each run a result (the platform's execution id and commit, the hash of the evaluation inputs, the evaluator's report, and the metrics file verbatim), marks each claim all of whose runs have results as verified, renders values.tex and tables/<key>.tex into .research/latex/{template}/ (each \airasval prints as a hyperlink pinned to the commit that wrote the record; table cells are always <row.run_id>.<column.ref_path>, so a label cannot be paired with another run's number), and commits what it wrote. It does not judge the result: a local verdict could be pushed past regardless, so the judgement is CI's, on the pushed commit. This is the only sanctioned way an experimental number enters the paper; never hand-write a tabular of experimental numbers. Honor the preregister-paper contract: the record is append-only, and a claim that missed its criterion is reported as a negative result, not rewritten.

   Run update_record last, after every other change to the record. The links in values.tex pin the commit that last wrote record.json, and the gate regenerates values.tex from that commit — so a render_chart or append_to_record made after update_record leaves values.tex pointing at a superseded commit and the gate reports values.tex differs from its regeneration. The tool is idempotent (a result already recorded is not appended twice), so re-running it after a late chart or append is the fix.