Human Compiler

Only the numbers. The lexer is pure code with no model involved, then one Jev request carries about 45 typed questions, roughly 30 Nouls, three Choices, four Scores and one Noul per paragraph. The model returns probabilities, not prose, and rules.ts compares each measurement to a threshold and emits the codes. That is why the same measurements always produce the same diagnostics, which the test suite covers.

Auto asks all nine profiles' questions in the same request, and the README's argument is that they run in parallel and cost only tokens, then it consumes only the chosen profile's answers. The pick comes from the dialect Choice and is reported as HC034 with the probability that drove it; below 0.50 it falls back to default, so a fuzzy input quietly gets the generic profile rather than an error.

The key never reaches the browser: the browser calls /api/compile, and the Worker verifies a Turnstile token, calls Jev, runs the rules and returns the report. Fronting it are a Turnstile widget and a per-IP rate limit of 20 compiles per minute, with the API key stored as a worker secret. Input is capped at 4,000 characters and 24 paragraph questions, so it is a memo linter, not a document pipeline.

The project says it outright at the bottom of the README: measurements are calibrated probabilities, thresholds are opinions. The nine lint profiles are opinions too, and teenager treating corporate phrases as errors or academic treating undeclared acronyms as errors are stylistic choices you can disagree with. The flags change the opinions further: -Wall lowers every threshold by 0.1, -Werror promotes warnings, and -O2 deletes paragraphs that add no information plus anything after "in conclusion".

Cloudflare Workers with static assets through @cloudflare/vite-plugin, React 19, Vite, TypeScript, @typesafe-ai/sdk for the model call, Cloudflare Turnstile in front of /api/compile with a per-IP rate limit binding, and Vitest for tests. Locally it is npm install, cp .dev.vars.example .dev.vars with your TYPESAFE_API_KEY in it, then npm run dev. The example .dev.vars uses Cloudflare's always-pass Turnstile test keys.

HC0xx are semantic thresholds, HC1xx structural, HC2xx lexical, and HC3xx mode-specific. The full table is generated from the rule catalogue onto the --help page inside the app rather than living in the README. Codes are stable across runs, which is why the same measurements always produce the same diagnostics: HC003 NullMeaningException for no recoverable information, HC201 for a detected phrase, HC034 for the dialect pick.

Yes. The rendered report prints a lexing line with milliseconds and paragraph, sentence and word counts, then an analyzing line with the model name, the question count and the token count, followed by severity out of ten and the exit code. In the README's own run that reads Lexing 0.4ms on 4 paragraphs and Analyzing 812ms on jev-1.13.0 with 45 questions and 1203 tokens.