AGENTS.md · diff
git:20260419.66fc11a to git:20260821.1f27c8f
7 added, 0 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
# Agent Instructions — unslop repo
Drop AI-isms. Keep the technical substance exact. Eat our own dog food: every doc, comment, and reply we ship in this repo follows the unslop rules.
## Workflow
1. Read [`CLAUDE.md`](./CLAUDE.md) for the full maintainer guide.
2. Edit only the SSOT files. Mirrors are auto-generated by CI.
3. Run `python3 -m pytest tests/unslop/` after any change to the Python package.
4. Run `bash hooks/install.sh` after editing any hook script and verify the `.unslop-active` flag updates correctly with `/unslop full` and `stop unslop`.
## Tone
- No "Great question!" openers. No "I'd be happy to help."
- No "delve", "tapestry", "testament", "leverage" (as filler), "robust" (as filler), "seamless", "holistic", "comprehensive" (as filler).
- No "It's important to note that". No "Generally speaking". No "In essence".
- No five-paragraph essay shape in commit messages or PR descriptions.
- Burstiness: short. Then one longer sentence that develops the point. Short again.
- Real uncertainty is fine ("I think", "probably") when honest. Performative softening isn't.
## Commits
Use `unslop-commit` style (Conventional Commits, body only when "why" isn't obvious from subject). See `skills/unslop-commit/SKILL.md`.
## Reviews
Use `unslop-review` style (file:line, severity prefix, observation, fix). See `skills/unslop-review/SKILL.md`.
## Tests are the contract
The `TestPreservation` suite enforces that humanization never modifies code, URLs, or headings. Don't disable it. If a test fails, the unslop is wrong, not the test.
+ ## Citation hygiene
+
+ When editing SSOT docs, verify benchmark attributions against `drafts/2026-08-detector-research/`:
+ - Liang ESL paper is **arXiv:2304.02819** (not 2306.04723 — that's Tulchinskii PHD).
+ - Jabarian & Imas (Booth 2025) tested **StealthGPT only** — not Turnitin, not twelve humanizers.
+ - **23.5×** few-shot style gain is **Jemama** (2509.24930), not Catch Me (2509.14543).
+
## Relationship to CLAUDE.md
This file is the short cross-IDE companion to `CLAUDE.md`. Keep the two aligned on policy; if one changes significantly, update the other in the same PR.