Skip to content

Trust & reputation

The marketplace earns trust through mechanisms layered on top of price discovery rather than upfront gatekeeping. Three are in place today.

Every provider posts a $50 USDC bond at enrollment, held while they operate. Serious misbehavior can trigger a manual ban with bond seizure; affected users are credited from a refund pool, with the treasury as backstop. The bond is released after a cooldown once a provider retires in good standing.

Providers accrue a reputation score from request outcomes (successes, failures, latency). Scores feed the sorted indices the coordinator uses for selection, so reliable providers are favored when prices are comparable. Stale providers (no recent heartbeat) are marked offline and excluded.

A small fraction of marketplace traffic (about 1%) is sampled as a hidden canary: a known prompt is substituted and the response is scored for deviation from the expected output. Runs that deviate beyond a threshold are flagged, and a sustained flag rate triggers manual review. This catches providers that misreport the model they serve.

Stronger guarantees — TEE attestation, content-addressed model registries, on-chain stake/slashing, and a privacy layer — are on the roadmap but intentionally deferred until marketplace traction justifies their complexity. Today’s stack is tokenizer-side counts, reputation, bonds, and canaries.