Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
21 lines (14 loc) · 2.15 KB

CONTRIBUTING.md

File metadata and controls

21 lines (14 loc) · 2.15 KB
Raw

Contributing

This project handles outside input the way its own tools say to: contributions arrive as returns to be sorted, not as patches merged on trust.

How a return is handled

Returns from other systems are divergence, not verification. That includes output from any AI system — mine or yours. Every return (issue, pull request, stress test, new corpus case) is sorted on receipt: folded (accepted, often reshaped) or declined, and either way the reason is written into the re-derivation log, Delta-Atlas-Tracer-Calibration.md. If your suggestion is declined, you get the reason in writing, in the log — not silence.

The rules

  1. The corpus is the floor. node tracer-corpus.js must pass before and after any detector change. When a case fails, the fix is to change the lexicon or rules — never the label. Found a real miss? The single strongest contribution is a new labeled corpus case with the input that produced it.
  2. AI-drafted stays candidate. Every change drafted by an AI — including the ones I draft with one — stays marked candidate until a human has done a cold read. No exceptions. That is the whole method.
  3. Structure over synonyms. A word added to a list is a fallback and gets labeled as one. A rule that catches the shape of a failure is a fix. If your change is a synonym, say so up front; it will be folded as fallback, not as the tripwire.
  4. Keep the constraints. Deterministic, client-side, zero dependencies, no model at runtime, lexicon versions displayed. A PR that adds a package or a network call will be declined no matter what it improves.
  5. Never commit *.local.js. Those are private seed-frame overlays — yours or anyone's. The .gitignore blocks them; don't work around it.

Practical bits

  • An issue with the input text, what you expected, and what the tool did is a complete contribution by itself.
  • Small pull requests, one concern each.
  • Vocabulary changes go to terms.enriched.json (the source of truth). Several pages embed copies of that data and have to be updated together, so flag any data edit clearly — a data PR that touches only one copy will be held until the copies agree.