cadgenbench / rules

# CADGenBench: prose audit

Auditable rules for prose on public-facing surfaces (READMEs,
`docs/`, dataset cards, Space app text). Process docs under
`space-setup/` are exempt.

Style basics (em-dashes, AI tells, lowercase) live in
[style.mdc](./style.mdc), not duplicated here.

## adhere to HF benchmark conventions
- Default to the shape established by other HF benchmark surfaces.
  When something isn't otherwise pinned, do what they do.
- Reference Spaces and repos:
  - [adyen/DABstep](https://huggingface.co/spaces/adyen/DABstep)
    (leaderboard Space text, validation guidelines)
  - [Open LLM Leaderboard](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
    (About / FAQ tone)
  - [huggingface/lighteval](https://github.com/huggingface/lighteval),
    [bigcode-project/bigcodebench](https://github.com/bigcode-project/bigcodebench)
    (benchmark README + docs/ layout)
- Short, concise, no fluff, user-friendly: document what the typical
  visitor of each repo / dataset / Space will do, and nothing extra.
- Order sections so the most common task scans first. Code repo:
  install + run, then internals. Dataset card: load + columns, then
  provenance. Space: submit + browse, then About / FAQ.
- Cut hedging, throat-clearing, bullet lists that exist to look
  thorough, repeated bold/italic emphasis, paragraphs that restate
  the section heading.

## describe what is, not what was
- Public surfaces describe the current contract. No change-log
  framing, no apology framing, no "previously / we used to" prose.
- Greppable tells (case-insensitive):
  `previously`, `used to`, `historically`, `originally`, `formerly`,
  `in the past`, `no longer`, `we now`, `we've moved`,
  `migrated from`, `removed`, `dropped`, `deprecated`.
- Release-history files (`CHANGELOG.md`, release notes) are exempt.
  Everything else: rewrite to describe the current state.

## don't define by absence
- Describe what something is, not what it isn't.
- Greppable tells (case-insensitive):
  `no separate`, `no need to`, `there is no`, `there's no`,
  `you don't have to`, `without having to`, `unlike`.
- Rewrite to the positive form. "Submissions land via the form" beats
  "no separate submission API to fill in".

## no false contrast
- "X not Y" / "X rather than Y" is only legitimate when Y is a real
  alternative a reader would genuinely consider (e.g. "the eval runs
  on the Space, not locally"). If Y is a strawman or an internal
  historical choice, cut the contrast and state the positive.
- Greppable tells (manual triage, many legitimate hits):
  ` not `, `rather than`, `instead of`, `as opposed to`.

## scope
- In scope:
  `cadgenbench/README.md`, `cadgenbench/docs/**/*.md`,
  `AI4Engineering/**/*.py` user-facing strings,
  `AI4Engineering/README.md`, dataset-card READMEs in
  `cadgenbench-data/`, `cadgenbench-data-gt/`,
  `cadgenbench-submissions/`.
- Out of scope:
  `space-setup/**` (process docs), `CHANGELOG.md`, release notes,
  test fixtures, code comments explaining historical decisions.

## audit recipe
- Run each greppable pattern above against the in-scope paths.
- For each hit: rewrite to a positive present-tense statement of the
  current contract, or confirm the contrast is genuinely useful to
  the reader. False-contrast hits are the largest bucket; most
  legitimate ` not ` uses survive review.