Skip to content

docs(eval): record forward-only rescore posture in spec 9.6 - #13

Merged
hsskey merged 1 commit into
mainfrom
fm/tokenloom-eval-c5-impl
Sep 17, 2026
Merged

hsskey merged 1 commit into
mainfrom
fm/tokenloom-eval-c5-impl

Conversation

@hsskey

@hsskey hsskey commented Sep 17, 2026

Copy link
Copy Markdown
Owner

Why

The rescoring contract in spec 9.6 did not state whether the historical one-shot rows can be rescored. A reader could not tell whether the deferred eval rescore command would ever ingest a recovered runs/out/ raw output. This records the settled posture so no recovery/ingestion path is expected.

What changed

Added one bullet to spec 9.6 (Reports and cost guard). It records that:

  • a rescore reads only the row's own embedded artifact.output, so rescoring is forward-only, with no path that ingests an external runs/out/ raw file back into a row;
  • the original one-shot raw outputs runs/out/2026-09-03 and runs/out/2026-09-05 are confirmed absent and unrecoverable, so the historical one-shot rows from those two dates are permanently non-rescorable;
  • a row written before artifact.output embedding is refused by the ordinary rule with no date or era special case, the same way section 9.5 treats a missing artifact as incomparable.

Documentation only. No code, no CLI, no snapshot contract, and no committed evidence changed. The eval rescore command remains deferred and unimplemented.

How I checked

pnpm verify --gate encoding

PASS encoding (crlf: 0, doubleEncoded: 0, nul: 0). No gate reads this prose as a threshold or lint target. Confirmed the edit touches only the spec 9.6 rescore bullet and leaves the section 9.5 S3-max and trajectory-selector prose intact.

What I left alone

No weakened gate, no vitest -u, no rewritten committed evidence, no refactor. Did not add or spec a rescore command, and did not touch the C2 changes (exact trajectory selector, S3 max gate) merged in #12. Spends no money and changes no public CLI or snapshot contract.

The one-shot raw outputs runs/out/2026-09-03 and runs/out/2026-09-05
are confirmed absent and unrecoverable, so the historical one-shot rows
from those dates are permanently non-rescorable. State that a rescore
reads only the row's own embedded artifact.output, that there is no
external runs/out/ ingestion path, and that a pre-embedding row is
refused by the ordinary rule with no date or era special case.
@hsskey
hsskey merged commit 624eb43 into main Sep 17, 2026
2 checks passed
@hsskey
hsskey deleted the fm/tokenloom-eval-c5-impl branch September 17, 2026 23:26
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant