What a verdict costs

Every figure on this page was measured on 2026-09-05 from public sources, and every one of them can be recomputed by anyone. The two claims below are the two that have settled on the current board, and the auditor finds no failed check on either. Sui gas comes from the transaction effects on a public JSON-RPC node. Walrus storage comes from the register transaction that paid for each blob. Model tokens come from the sealed run bundles, which record every attempt including the repairs. Research calls come from the sealed tool transcripts. Nothing on this page is read from the operator's database, and the script in Recompute it rebuilds the whole table from the same public endpoints.

One line is an estimate rather than a measurement, and it says so. Everything else is a number that exists on a chain, in a blob or on a published price list.

Testnet SUI and testnet WAL have no market value. The dollar columns convert native units at the mainnet rate of the day, so they read as what the same work would cost if this deployment ran on mainnet.

What was measured

Both juries seated five seats across two model families under the degraded diversity rule the certificate records. Neither claim split, so neither opened a debate. Attempt 1 of the Bitcoin claim was voided on an invalid schema, and its cost appears in The entire run.

Unit prices

GonkaRouter charges one rate for every model it serves and bills prompt and completion tokens the same, so the three families on this deployment share a price. Firecrawl's credit price depends on the plan, and the plan a deployment is on is not public. The tables below use the Hobby rate, which is the smallest paid plan. On the Standard rate the research line falls to about a quarter of what they show.

One verification

A verification is five juror seats, each researching the claim on its own, sealing its work to Walrus, pinning a hash on Sui, committing a hidden vote and then revealing it. Four components carry the money. Sui gas pays for the protocol transactions and for the two transactions every Walrus write needs. Walrus pays for the bytes. GonkaRouter pays for the model turns. Firecrawl pays for the searches and page opens the engine runs on the jury's behalf.

Sui gas

Every row is the sum of the transactions named, with gas taken as computation plus storage less the storage rebate, which is what the sender actually paid.

Three things in that table are worth a second look.

The vote commitment is negative. Committing consumes objects the run approval created, and Sui refunds the storage they held, so the seat key ends the transaction with more SUI than it started with. The reveal is nearly free for the same reason.

The committee draw is the single most expensive protocol transaction, ahead of the finalize that mints the certificate. It reads Sui's shared randomness object and mints five seats in one call.

Walrus writes cost more gas than the protocol itself does. Registering and certifying the blobs is the larger half of the bill on both claims.

Walrus storage

Each write is one Sui object, so a blob written twice is paid for twice. The raw and canonical copies of an evidence artifact are two writes even when the bytes are identical.

Blob size does not move this number. Walrus erasure codes every blob across a thousand shards, and the coding overhead dominates at these sizes, so the shortest statement and the largest run bundle reserve exactly the same encoded size and round to the same number of storage units. What a claim pays Walrus is set by how many blobs it writes, not by how large they are.

The register transaction splits the exact WAL it is about to spend and destroys the remainders, so the two paid figures are readable on chain rather than inferred. Recomputing them today with the Walrus SDK's storageCost helper gives a little less, because the price per storage unit is a live value on the Walrus system object and it has moved since the write.

GonkaRouter inference

The bundle records every attempt a seat made, not only the one that was accepted. Hedges that hit a provider error and repair turns that fixed a malformed answer are model time the operator paid for, so they are counted here.

A priced call is one that came back with a usage record. The rest were provider errors that returned nothing, and GonkaRouter's public receipt for a failed call reports no tokens. The Great Wall claim spent much more model time than the Bitcoin claim because its seats took more turns to reach a valid answer. That is the honest reason the two claims differ, and it is why a single verification has no fixed inference price.

Web research

Search and open are engine-executed. The engine keeps one cache per claim, so the first seat to run a query pays for it and the other four read the same result.

Both claims landed on the same credit count. A failed open returned no page, and Firecrawl's published table charges per page returned, so the failures on the Bitcoin claim are counted apart and not billed. That is an assumption about Firecrawl's billing rather than something a receipt proves, and the whole question is worth a few credits either way.

The four components together

Inference is by far the cheapest thing a verification does. Five independent AI jurors researching a claim and voting on it cost a small fraction of a cent, and the settlement around them costs orders of magnitude more. Most of that settlement bill is not the protocol itself. It is the gas for the two Sui transactions that every Walrus write needs.

A second round

Neither settled claim split, so neither opened a debate, and no claim on the current board has been through one. The money cost of a second round is therefore not yet measured. What the code adds is countable from How a verdict happens.

Round two runs no searches and opens no pages, so it adds no research credits. On the shape above a split claim would cost roughly twice a settled one, with the extra weighted the same way, mostly Walrus writes rather than inference.

The entire run

The public board serves what the operator's record holds today. It served the four claim records below when this page was written, on 2026-09-05 at 10:37Z, and new verifications keep landing on it. Verifications that ran before the operator reset that record on 2026-09-05 are no longer served by the API, so the model tokens and research calls they spent cannot be recovered. Their gas is not lost, because every transaction these keys ever sent is on chain, and the sweep below reaches back to the first one a public node will still return.

Three groups of addresses pay for this deployment.

Sweeping all of them on 2026-09-05 gives the whole on-chain bill.

That total is a floor, not a ceiling. The public JSON-RPC endpoint is a pool of nodes and they do not all keep the same history, so a request for an old transaction sometimes comes back empty. The sweep counts what it could not read rather than guessing at it, and those transactions carry gas that is not in the figure above. A node with full history would report more.

Walrus is charged per blob rather than per address, because the WAL a writer lane spends was transferred to it by the operator first, and counting both movements would count the same money twice.

Here is what the board itself has cost. The voided attempt is a real cost, and so is the verification that settled without reaching a verdict.

A voided attempt shows no model tokens because its seats never revealed. Their bundles are still sealed on Walrus, so the tokens they spent are real but not public, and the page counts only what a reader can check. The gas and the blobs of a void are public, and they are counted.

Five transactions put the contracts on chain and four of them are readable today. They are already inside the operator's total above rather than additional to it.

Three more amounts sit on chain without being consumed, so none of them is a cost of running the protocol.

What a claim would need to cost

Two numbers set a price. The variable cost is what one verification consumes, which the tables above measure. The fixed cost is hosting, the one estimate on this page.

Spreading the hosting over a month of claims gives a break-even price. The last column compares it with the escrow the operator locks into each claim today.

Those rows price a verification that settles on its first attempt. One that has to be relaunched pays for the attempt that failed as well, and the board's own record of that is in The entire run.

What the protocol charges today

Nothing is charged to whoever submits a claim, so there is no revenue to set against the table above. The escrow is the operator's own money, and the fee ticket the finalize mints is paid to the operator's own treasury. The fee is a parameter rather than a constant, set on chain with set_treasury_policy under the ceiling above, so raising it changes how an escrow is split and not what a submitter pays. Covering the break-even price would mean charging the submitter, which this deployment does not do.

The gap is wide but the shape of it is encouraging. The expensive parts are the settlement receipts, which shrink with batching and with fewer, larger blobs, and the research credits, which fall on a bigger plan. The jury itself, the part that does the actual work, is close to free.

Recompute it

The tool that produced every table above is in the repository and reads only public sources.

# one claim, native units only
pnpm cost:claim 0xd86526512ffcc78c3460064b46b56252d5d2d3adc69aedc24ad9c8325c81ac8c

# the same claim in dollars, at the rates this page used
pnpm cost:claim 0xd86526512ffcc78c3460064b46b56252d5d2d3adc69aedc24ad9c8325c81ac8c \
  --sui-usd 0.7753 --wal-usd 0.02693 \
  --gonka-usd-per-mtoken 0.0012 --firecrawl-usd-per-credit 0.0038

# every claim on the board
pnpm cost:claim --board

# the board plus every address that paid for it
pnpm cost:claim --run-total

The auditor that proves the claims themselves is a separate tool, and the audit guide walks it end to end.

pnpm ov audit 0xd86526512ffcc78c3460064b46b56252d5d2d3adc69aedc24ad9c8325c81ac8c

Both tools read Sui through the JSON-RPC fallback in config/release.testnet.json, which is a public endpoint that needs no key.

Five caveats belong with the numbers.

Prices move. SUI, WAL and the Walrus price per storage unit are all live values, so a rerun will not match this page to the last digit. The native units will match, because they are on chain.

Testnet tokens are free to obtain and worth nothing. The dollar columns are a mainnet equivalent, not a bill anyone paid.

Gas on mainnet uses the same units and the same rules, so the MIST figures carry over. The reference gas price does not, and a busy network charges more.

The board moves. Rerunning the board or the run total will find a different set of claims, because verifications keep landing and the operator's record can be reset.

The Firecrawl plan of this deployment is not public, so the research line is the only component priced from a plan the page had to choose rather than read.