Skip to content

feat: block venue-ambition leakage in publication-facing prose - #9

Merged
mxuanvan02 merged 1 commit into
mainfrom
feat/venue-ambition-gate
Sep 6, 2026
Merged

mxuanvan02 merged 1 commit into
mainfrom
feat/venue-ambition-gate

Conversation

@mxuanvan02

Copy link
Copy Markdown
Owner

What

Adds venue_ambition_leak as a blocking class in the internal-register gate. The publication target -- venue tier, impact metric, intent to submit, or an anticipated referee reaction -- is a planning decision between the author and their collaborators. Inside publication-facing prose it is unusable: a reader of the published paper cannot act on any of it.

Licensed by genre, not by wording

The same sentence is correct in a cover letter and blocking in a manuscript, so the class is licensed through GENRE_LICENSED (cover_letter, revision_notes) rather than through phrasing. Both directions are pinned by test: a gate that only ever fires proves nothing about its licensing.

Calibration was the hard part

The first pattern set produced six false positives on correct academic prose:

Correct prose Why it fired
interquartile range bare quartile
the first quartile falls at 0.79 bare quartile
Revenue in Q1 of the observation window venue inside Re**venue**, and Q1 as a fiscal quarter
retrieved from Scopus and Web of Science index named as a search source in Methods
the impact of annotator disagreement bare impact

Fixed by requiring co-occurrence with a venue/submission term and by adding word boundaries. All six are kept as permanent tests so the same false positive cannot return at the next pattern edit.

Verification

Four mutations, each restored green afterwards:

  • remove the code from BLOCKING -> caught by validator
  • unlicense the class for cover letters -> caught by validator
  • delete the Vietnamese referee pattern -> caught by validator
  • delete the rubric blocking bullet -> caught by repo tests
  • restore the Revenue false positive -> caught by validator

The rubric mutation initially passed: assertIn("venue_ambition_leak", rubric) stayed green because the code also appears in the explanatory paragraph below the list. The assertion now pins the bullet form for all 23 codes. That pitfall, and the Revenue substring lesson, are recorded in references/skill-repository-maintenance.md.

Suite

validator          rc=0
repo/schema tests  25 OK
internal_register  38 OK  (was 29)
process_logic      13 OK
vi_ai_pattern      30 OK

The publication target is a decision made between the author and their
collaborators, not a property of the work. A manuscript that argues its own
venue tier, states an intent to submit, or anticipates a referee reaction has
imported the planning conversation into the artifact, and no reader of the
published paper can act on any of it.

This is a strategy leak rather than a wording defect, so the class is licensed
by genre instead of by phrasing: the same sentence is correct in a cover letter
and blocking in a manuscript. Both directions are pinned by test, because a
gate that only ever fires proves nothing about its licensing.

Calibration mattered more than detection here. The first pattern set produced
six false positives on correct prose -- interquartile ranges, fiscal quarters,
Scopus named as a search source in Methods, and the substring `venue` inside
`Revenue`. Co-occurrence with a venue/submission term plus word boundaries
removed all six with no loss of recall; the cases are kept as permanent tests.

Verified by mutation: removing the code from the blocking set, unlicensing the
cover-letter genre, deleting the Vietnamese referee pattern, deleting the
rubric bullet, and restoring the Revenue false positive each turn the suite
red, and each is restored green afterwards.
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

@mxuanvan02
mxuanvan02 merged commit 773e218 into main Sep 6, 2026
2 checks passed
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