What it is
Classifies text-bearing PDF pages into IRS form and page-kind candidates with a confidence gate, for teams triaging tax documents. It is document triage, not tax advice, and scanned pages need OCR first.
How it uses Jev
Each page's text is sent to Jev as one request over 7 page kinds and 230 form options (plus not_in_this_list); five corporate/foreign forms trigger a second sub-form question. Jev returns probabilities, and formConfidence (the minimum over steps) gates whether the caller acts or falls back.
Primitives:choice
Technique worth stealing
Gate on the minimum confidence across steps; fall back below the gate instead of trusting a low-confidence label.
Try it
pnpm install; pnpm typecheck; pnpm test. Needs Node 20+, poppler, and TYPESAFE_API_KEY.
Evidence
Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.
- Jev-centricyes0.95
- Shows a System One patternunclear0.53
- Handles uncertaintyno0.15
- Measuredyes0.82
- Runnableyes0.87
- Worth recommendingunclear0.46
- Model replicano0.05
- Problem scopescore on a 0–2 scale1.04
- About Jevyes0.99
Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.