Blog.

Short notes you can cite: how we measure over-refusal and test guardrails. Lawful research only.

How we benchmark over-refusal with research presets

Lawful fiction edge-cases, with/without preset, false-refusal rate as the metric. Method + checklist.

6 min2026-10-18

A decision rule for developers: API for harnesses and refusal benchmarks, chat for personas. Workflow + code.

5 min2026-10-18

Preset vs system prompt vs fine-tune: per-request scope, what relaxes on lawful creative, what stays refused.

7 min2026-10-20

Your app in the middle, RawLine as adversarial generator, lawful edge-cases in a loop. Metrics + harden cycle.

6 min2026-10-20

GEO + SEO template: definition first, method, raw rates, FAQ schema, llms.txt entry. Copy the frame.

5 min2026-10-20

For educational and research purposes only. Lawful use only.