[Nikit117]

The row schema and the shared steps are already documented, in
docs/extractbench.md#submitting-results
and README#submitting-results.
What those pages do not cover is the other half of a submission: how a row
gets checked, what that check costs a reviewer, and what a submitter can do to
make it cheap.

This adds docs/verifying-a-submission.md, written from the submitter's side:

• what checking a row costs: scoring the submitted predictions takes minutes,
  running the pipeline over the corpus is the expensive part
• publish the per-document predictions and scores, so a reviewer can recompute
  the score without running anything and without spending an API key
• state the endpoint, every service the run needs (not only the model), the
  configuration switches and the cost derivation before the reviewer has to ask
• pin the revision by a tag as well as by a commit, so a force-push cannot move
  the numbers a row was produced from
• offer a one-command subset run that prints the published per-document score
  next to each document that just ran
• what actually caused a round trip on our own submission, and a checklist for
  submitters

The running example is our own submission (#84), which is why the "what caused
a round trip" list is specific rather than hypothetical. I am happy to cut that
section if a more general list is preferred, and to fold the checklist into
docs/extractbench.md instead of keeping a separate page.

No change to the inclusion criteria, the schema or the scoring - this is a
submission-side guide only.

[Nikit117]

Closing this for now. The submission side is documented properly in
docs/extractbench.md, and this page turns into an opinionated guide that the
repository did not ask for - and pushing a docs pull request into a repository
where my other pull request is already open is not the right move.

The file stays in my fork (docs/verifying-a-submission, commit 03c1440) if
it is ever useful.