Panel starts as free bot protection for normal sites. The platform loop is the moat: verification traffic creates abuse telemetry, the edge model improves, and paid teams can layer panel-data, rater workflows, and agent-output feedback on top.
operator stack panel surface raters
───────────── ───────────── ──────
agent / model ──┐
│ emitMedia()
modal worker ───┤ emitProcessOutput() ┌── L1 · taste captcha ──┐
├─────────────────► panel ──┤ (any visitor, gated) │──┐
comfyui ──────┤ panel-sdk ingest │ │ │
replicate ─────┤ hmac · batch api ├── L2 · agent rating ───┤ │ judgments
elevenlabs ────┘ │ (trusted raters) │ ├──────────►
│ │ │
└── L3 · expert review ──┘ │
│
◄──── signed preference rows ──────────┘
◄──── rater balance ledger
(paid-train · coming)
three layers of human signal, one shared rater pool. trust scores propagate up — raters that handle L1 cleanly get promoted to L2, L2 performers with verified credentials get gated into L3.
panel-sdk is the operator side. it handles hmac signing, batching, retries, and the ingest contract. adapters wrap common generation surfaces so operators don't have to wire a webhook from scratch.
emitRaw(units) for anything else.raters earn balance per accepted unit. balance converts to inference credits and finetune runs against the dataset they helped label. the operator that paid for the rating uses the resulting model on the same surface that produced the data. no broker in between.
paid-train is private alpha. request access.