panel — the agent-output demo
panel is a captcha. what makes it different: instead of clicking traffic lights to train someone's self-driving stack, your visitors judge your agent's actual output. skill diffs, tool-call steps, pairwise trace comparisons. you keep the labeled dataset. captcha-tier cost, RLHF-grade signal.
what an anonymous visitor sees. taste calls, media judgments, simple comparisons. low friction, broad participation. this is what gates a signup form when you embed the widget.
this is the wedge. trust-tier raters (your dogfooders, your power users, your team) judge real agent traces from your stack: a skill diff before/after, a tool-call step in context, a pairwise of two reply candidates. every judgment is a labeled example you own. no scale.ai contract, no surge invoice, no labeled-data vendor in the loop.
that's how a visitor sees it during signup — pill on the form, public-pool unit on click, returns a verified token. you keep the judgment. the operator console shows you the unit + the answer + the rater's trust tier.
<div data-panel-sitekey="pk_demo_a" data-panel-mode="pill" />