Triage: catalogue EvoRiskBench, PAGE and STEER-Bench (#196, #194, #198)

[emmanuelgjr]

Three rows for evals/EXTERNAL_BENCHMARKS.md, from the watcher triage of the issues below, plus a precedent note in docs/TRIAGE_RULES.md rule 3.

Closes #196
Closes #194
Closes #198

The rows

Benchmark	Paper	Scale, as published	Release	Disclosure in row	OWASP entry (DRAFT)
EvoRiskBench	arXiv:2610.03153	450 adversarial tasks, 6 scenarios, 9 model-harness configs	not released: the paper says it will be released "after completion of artifact safety and reproducibility checks"	—	ASI02 · ASI05
PAGE	arXiv:2610.03434	not stated in the abstract	not released: the arXiv record links none	the authors' own guardrail (Persona Guardrail) is evaluated on it	LLM01 · LLM03
STEER-Bench	arXiv:2610.03089	101 tasks, 9 domains	not released: the arXiv record links none	the authors' own defense (COBRA) is evaluated on it	LLM01 · ASI01

Every figure and quotation is taken from the paper's abstract, as the catalogue's header requires. Nothing has been re-run.

The precedent (maintainer ruling, 2026-10-09)

This PR adds the first catalogue rows for benchmarks that are a secondary contribution of a defense paper. Rule 3 now says:

• a defense paper that also introduces a named benchmark gets a row for the benchmark;
• the defense itself is noted;
• the row names the authors' own defense evaluated on it;
• a defense paper with no named benchmark is noted and closed, as before.

C4

DRAFT — SME review required. The OWASP entry column is a reading of each benchmark's stated scope, not a reviewed mapping, under the catalogue's existing DRAFT banner.

Checks (local)

• validate.js: 0 errors / 92 warnings / 329 passed, the same before and after.
• test:scripts: 137/137.
• stats:check: current.
• generate.js: no drift.
• markdownlint: no new findings in either file. The 16 pre-existing MD060 findings come from a newer local linter than CI's v16.

The link check doesn't run on PRs. The three arXiv URLs resolved when checked today.

🤖 Generated with Claude Code