humanoid.js scores browser-agent click traces

This is a small but timely eval: browser agents are improving at task completion, but their low-level input traces still often look synthetic.

• –The strongest angle is transparency: every rule is readable client-side, so teams can compare agent strategies without sending data to a hosted detector.
• –Signals like pressure variance, contact geometry, jitter, path efficiency, sample timing, WebDriver markers, and UA mismatches map well to the kinds of rough edges automation stacks expose.
• –It should be treated as a smoke test, not proof of stealth or compliance; passing a toy scorer does not mean passing Cloudflare, DataDome, or PerimeterX.
• –For defensive QA, it gives agent developers a concrete way to catch unnatural clicks and drags before they hit real workflows.