Your buyers pay for Opus.
Prove they're getting it.
Some of your sellers advertise Opus and quietly serve something cheaper — and when quality drops, the buyer blames your marketplace, not the seller. Litmus is the neutral layer that catches them: independent attestors test what each seller is actually serving and post the verdict on-chain.
You can't verify your own sellers. And buyers know it.
Sellers win the cheapest-first race by quietly swapping the model. A request for Opus comes back from Llama. Your buyer's product quality drops — and they blame your marketplace, not the seller. You profit from the cheapest offer, so self-verification isn't trust. It's marketing.
We can already show you who's lying.
Our probes run as ordinary buyer traffic — indistinguishable from a real request, so a cheating seller can't dodge them. Point us at your marketplace and we'll hand you a report: every offer, the model it claims, the model it's actually serving, with on-chain evidence anyone can replay.
No integration required to get the report. You'll know within a day.
Catch the swap before your buyers do.
Litmus is a verification protocol, not a service you have to trust. Independent attestors stake real money, probe your sellers, and publish a score on-chain. You just read it — and the cheaters get caught while the buyer's still happy.
Not your job to be trusted.
You're conflicted — you earn on the cheapest seller. Litmus is the disinterested third party. Buyers believe a score you didn't produce.
Every score is auditable.
Scores, evidence hashes, and slashing events settle on Base. A buyer — or you — can replay any verdict. It's proof, not a press release.
You run nothing.
No probes, no attestors, no settlement to operate. Litmus handles the testing economy. You add one field to your router.
Building it yourself doesn't solve the problem.
In-house verification
- You're the conflicted party — buyers discount your own "verified" stamp
- Months of probe engineering, on-chain settlement, and attestor coordination
- You still have to be trusted not to protect your top-earning sellers
- No shared standard — every marketplace reinvents it differently
Litmus
- Neutral by construction — the protocol, not you, issues the verdict
- Auditable on-chain — buyers replay the evidence themselves
- Live in an afternoon — one read on your routing path
- A shared badge buyers learn to look for across the market
One read on your routing path.
Verification runs async, entirely off your hot path — it never touches your latency or your margins. Read a seller's score from an edge KV cache in sub-millisecond and fold it into the sort you already run. Feature-flag it. Ship it behind a toggle. No crypto knowledge required to read a number.
// litmus score: 0–10000 bps, edge-cached const score = await litmus.get(offer.id) // KV, <1ms // fold into your existing cheapest-first sort const effectiveCost = estimatedCost / (1 + score / 10000 * WEIGHT) // WEIGHT = 0 → behavior unchanged // WEIGHT = 1 → prefer a verified seller // over a 50%-score seller at 1.5x price
The badge buyers start filtering for.
Surface the score next to price. Honest sellers earn a premium; cheaters get routed last. You become the marketplace buyers recommend — because it's the one that proves what it sells.
Why the score can't be gamed.
Attestors put up USDC
Verifiers stake real money to participate. A dishonest verdict costs them their stake.
Probes look like traffic
Tests route through your normal buyer path. A seller can't serve the real model only when watched.
Liars get slashed
Caught downgrading? The seller's stake is slashed on-chain and the offer drops in routing.
Anyone can dispute
Every result has a challenge window. Re-run the probes, post contradicting evidence, win the slash.
OpenAI-compatible from day one. If you route inference, Litmus reads in one call.
See who's lying on your marketplace.
Prefer to talk first? Book a 15-min call →