a captcha you don't hate. visitors prove they're human by judging one piece of agent output. the judgment goes back to the operator as preference data. recaptcha trains google's self-driving cars on your visitors. panel trains your own systems.
no traffic-light squares. one taste judgment — pick the better headline, flag the off-sync dub, rank three replies. takes 3 seconds. the answer is a row of preference data you own.
public captcha pool is taste, sarcasm, dub-sync, voice-naturalness, perception. things a frontier LLM cannot reliably solve in 2026. technical judgments (code, traces, hallucination calls) never touch anonymous raters — they flow through a paid trust-tier pipeline. this is the wedge. see D12.
behavioral floor (mouse, dwell, focus entropy). engagement window. honeypot units where the obvious-LLM-answer is wrong by design. opaque scoring — the token issues unconditionally, the probability resolves hours later once humans agree. bots can't tight-loop the verifier.
scrubber-proxy sits in front of the unit pool. GDPR / KVKK / HIPAA-aware redaction before any judgment leaves your stack. use the proxy on its own if you want.
<script src="https://panel.goku.codes/sdk.js" async></script>
<div data-panel data-operator-key="op_xxx"></div>
// on success: a token. verify server-side.
fetch('https://panel.goku.codes/api/verify', {
method: 'POST',
headers: { 'x-operator-key': process.env.PANEL_KEY },
body: JSON.stringify({ token })
})the iframe widget at /embed works today. the script tag is the alpha shape. PoC, not GA.
fake signup form. judge one unit, get through.
postMessage handshake. embeddable today.
recent judgments, trust score, agreement.
operator key, embed snippet, traffic, dataset preview.
panel · proof of concept · github