Verifiable AI inference on FLOP: proof of which model answered

Most hosted AI APIs ask you to trust that the model you requested is the one that answered. FLOP is built to give the buyer evidence. A session request pins a model-weight hash, and the miner must leave a proof that "the requested inference was performed on the requested model," in the words of the whitepaper. This use case is for builders who need that evidence, such as an agent that has to show which open-weight model produced its output, or a buyer comparing providers on more than price.

Status, November 10, 2026: the proof format is specified and the devnet runs sessions, but the default SOFT tier's "end-to-end settlement and dispute lane is still being specified," per Flop's PoI doc. The testnet is not live. See the compute use cases hub.

What a FLOP proof shows, and what it doesn't

The PoI doc says each session leaves "activation commitments and a co-signed transcript." During the session the model commits fingerprints of its internal state, "TOPLOC commitments, 258 bytes per 32 tokens," and every turn becomes a co-signed leaf under a Merkle root that the chain stores at settlement.

The same doc is clear about the limit: "PoI establishes that the registered work was performed as specified. It does not establish that a model's answer is correct, safe, or high quality." You get evidence of which model ran and that the work was done, not a guarantee that the output is right.

How the checks work

The whitepaper combines four layers and relies on none alone: optional hardware attestation (TEE), the TOPLOC activation fingerprints, randomised re-execution "against the original prompt by bonded checkers," and miner stake at risk. The challenge game doc adds the numbers:

Mechanism	What the docs say
Audit lottery	Random turns are selected each epoch, "drawn across the work by value"
Audit rate	2.5%, so a single cheated slice slips through about 97.5% of the time
Who can challenge	The paying agent, or any active validator
Challenger bond	100 FLOP, forfeited if the dispute is frivolous
Miner response time	Two hours to produce the disputed turn's evidence; "Silence defaults to a fraud verdict"
Penalty for fraud	100% of stake slashed, ejection and blacklist; the agent's escrow is refunded
Challenge window	About 7 days, "under review to shorten toward hours"

The verification FAQ says re-execution "replays the original user prompt," so "an audit is indistinguishable from ordinary traffic." Our proof of useful inference explainer walks through the full design.

How a builder would use it

1. Pin the model. Choose the exact registered model. The direct session guide says a miner "must serve the exact registered model hash, measured root, precision and decode policy," and ambiguous versions "are refused before escrowing."
2. Know the quantisation. Flop's demand FAQ says the check is calibrated to the model, version and quantisation, "so a registered 4-bit deployment is verified as a 4-bit deployment."
3. Keep the receipts. The SDK's runDirectSession "verifies every receipt before accepting output or signing payment." Store the channel ID and settlement transaction with each output.
4. Challenge when it's wrong. An agent that suspects the work can dispute its own session instead of paying, as described in FLOP for AI agents.

Does verified inference earn more airdrop?

No extra credit is described. The draft airdrop page shares the agent pool "pro rata to compute purchased in settled sessions." Every testnet session is verified the same way, so a buyer who keeps receipts earns the same as one who doesn't for the same compute. On the miner side, only "inference that passes verification" counts. Devnet earns nothing. Read the airdrop guide for the full rules.

Honest limits

• The SOFT dispute path is unfinished, and the attested settlement path is the one currently wired.
• Verification proves the work, not the answer's quality.
• Disputes about off-chain deliverables are out of scope. The whitepaper says "counterparties choose their own adjudication."
• The 1% audit earmark from session fees is "documented now and wired later," per the challenge game doc.

FAQ

Can FLOP prove which AI model answered a request?

That is its design goal. A session pins a model-weight hash, and the miner leaves TOPLOC activation commitments and a co-signed transcript that validators record and challengers can re-check. Flop says the proof shows the work was done as specified, not that the answer is correct.

How often does FLOP audit inference?

Flop's challenge game doc gives a 2.5% audit rate, drawn by a random lottery across the work by value. Paying agents and active validators can also challenge any session, with a 100 FLOP bond for discretionary challenges.

Is FLOP verification live?

Partly. Sessions run on devnet, but Flop says the SOFT tier's end-to-end settlement and dispute lane is still being specified. The testnet, where verified work first counts, had no launch date as of November 10, 2026.