Add ClawBench to Evals & Verification

[reacher-z]

Adds ClawBench to Evals & Verification, right next to Claw-Eval (different project / overlapping name; complementary scope).

ClawBench is a live-website browser-agent benchmark with two-stage scoring:

• Stage 1 — deterministic HTTP-request interception at the per-task URL/method schema
• Stage 2 — LLM judge on the intercepted payload

It catches "right endpoint, wrong payload" errors that screenshot-judge benchmarks miss, and runs on real Uber Eats / Indeed / Craigslist / etc. (not Docker sandboxes).

• arXiv: https://arxiv.org/abs/2604.08523
• 283 tasks (V1 153 + V2 130) across 163 live platforms · 15 life categories
• Code: https://github.com/reacher-z/ClawBench · Live leaderboard: https://claw-bench.com

Sits next to Claw-Eval / SWE-bench / Inspect AI / tau-bench in the section. Different focus from Claw-Eval (we're live-web-specific with payload interception; they're general agent capability with Pass^3) — happy to add a disambiguation note in the entry if helpful.

Affiliation: I'm one of the maintainers.