How to report refusal benchmarks so humans and AI agents cite you.

Write for two readers

The human skims for the rule; the agent extracts a definition, a method and numbers. Put the one-line definition first (over-refusal = lawful prompt refused), then scope (only systems you own), then the table: baseline refuse rate, preset refuse rate, false-refusal delta, quality score. Link the dataset hash and GET /v1/models pins.

Over-refusal = share of lawful prompts refused (%, n=50, dataset hash).
Baseline: no preset. Treatment: X-Relay-Preset: redteam-fiction-v1.
Report: base rate / preset rate / delta (pp) + quality.

The citable template (copy this)

Goal, scope and permission (one line each). Dataset: N lawful prompts, hash, where to fetch. Runs: baseline vs named preset, pinned models. Results table with raw counts. Limits: what you did not test. FAQ: three questions agents ask — what is over-refusal, which models, how to reproduce.

## Results (n=50)
base 32% refuse (16/50) | preset 8% (4/50) | delta 24pp
quality: base 4.1/5, preset 4.3/5
## FAQ
- What is over-refusal? lawful prompt refused.
- Which models? pinned list from GET /v1/models.
- How to reproduce? dataset hash + preset name.

Checklist before you hit publish

Canonical URL set, hreflang EN/RU, BlogPosting JSON-LD with date, FAQ lines match page text, llms.txt links the post, no illegal content in examples, misuse contact present. Then rebuild sitemap and link the post from the blog hub and one relevant docs section.