PAL 9000 — Practice And Learn
Practice-quiz site for CIS 9000: IT Strategy (Baruch College). Weekly scenario-question sets that mirror the exam's case-based reasoning, plus a spaced review queue for missed questions. A HAL 9000 backronym — but PAL is on your side.
Built on Groundwork: a grounded generate → curate → measure pipeline for LLM-authored practice questions.
How it works
- Students open a link — no account needed. Weekly sets of ~10 questions with immediate per-distractor feedback and self-explanation prompts. Missed questions come due again after 1, 3, then 7 days (SM-2-lite; ts-fsrs is the upgrade path). Weeks are grouped into two sets, each closing at a checkpoint (Test 1, then the final): finish a week and it's stamped with a short completion code tied to the name entered on the device, and you hand in the whole set's log once from the
/systems/ page instead of screenshotting every week. Progress lives in localStorage only, with self-serve backup/restore. The site is an installable PWA built for subway commutes: the full app shell precaches on first visit, sets keep working offline, and an interrupted quiz resumes where it left off.
- The instructor drops each week's materials into
content/sources/week-NN/ (gitignored — course/publisher content never enters git), runs npm run generate <week>, and curates the over-generated candidates into content/questions/week-NN.yaml. Accept/reject stats per prompt version are logged in docs/EVAL-LOG.md — the eval loop is the point.
Commands
npm run dev # local dev server
npm run build # build to _site/
npm test # vitest (schema lint + scheduler + built-site checks)
npm run generate 5 # generate week-05 candidates from content/sources/week-05/
npm run verify logs.txt # check pasted completion logs in bulk (--set N, --csv)
Generation runs locally only (Claude API; needs ANTHROPIC_API_KEY or an ant auth login profile). Sources: .md/.txt read as text, .pdf sent natively — export PPTX decks to PDF first.
Question lifecycle
candidate → curated to accepted / edited / rejected: <reason> (hallucination | leakage | trivia | ambiguous | duplicate). Only accepted/edited questions build into the site; the schema lint in tests/unit/bank.test.js enforces citations, per-distractor feedback, and self-explanation prompts.
Deploying
Netlify, gated: production builds run only when the commit subject contains [deploy] (see netlify.toml).
Status
Live at pal9000.netlify.app (August 2026): app, scheduler, pipeline, offline/PWA support, storage management, per-set checkpoints with name-bound completion codes, and tests are built and deployed. Question content is still hand-authored sample material (promptVersion: sample-v0) pending the first real generation run against fall course materials.
Completion codes are tamper-evidence, not proof — the salt ships in public source, and there's no backend to check against. The bar is "harder than doing the set"; the credit at stake is a small test bonus and stays honor-system. Names never leave the browser except when a student pastes their own log into the LMS.
Testing tip: to exercise the review queue without waiting a day, temporarily shrink INTERVALS_DAYS in src/assets/js/scheduler.js.
License
Code is MIT. The question content under content/ is not MIT — see content/LICENSE (all rights reserved; instructors interested in reuse, get in touch).
Full spec: docs/SPEC.md.
