hackathonBLI Legal Tech, Oct 2026

vouch

99.3% similar is still wrong

A citation check for AI-written research. Every figure an AI cites about a public company is checked against the SEC filing it quotes, and a wrong claim loses its stake.

Solidity
Chainlink CRE
Python
SEC EDGAR
99.3%

similar to the filing, with one figure changed. A 95% fuzzy match passes it.

10 / 10

doctored quotes slashed in one round, all of them 99.1% to 99.7% similar.

50 / 50

verdicts written by the CRE workflow on Robinhood Chain testnet: 18 cleared, 32 slashed.

The problem

Lawyers have been sanctioned for briefs citing cases an AI made up: 1,116 US court filings by June 2026, by one public count. Investment research is now drafted by the same models and fails the same way. An AI wrote “Depreciation for Q3 2026 was $1.37 billion” about Tesla; the 10-Q sentence it cited gives that figure for the quarter ended 30 June. FINRA Rule 2241 asks that facts in a research report rest on reliable information. A model cannot certify what it believes, but a firm can show that every number was checked.

  1. Post the claim
    One figure, the filing sentence behind it and the filing's accession number, hashed on-chain with a stake.
  2. Fetch and compare
    A Chainlink CRE workflow fetches the filing from SEC EDGAR and runs the deterministic check.
  3. Keep or lose the stake
    A wrong claim loses its stake to a readers' fund in the same transaction as the verdict.
  4. Trade only on what held
    A trading agent's desk refuses, on-chain, any trade citing a claim that failed or was never checked.

Key decisions

Exact figures, not a fuzzy match

Citation checkers usually accept a near-match, to survive the noise of turning a filing into text. The verifier carried over from an earlier project did, at 95% similarity. In a financial filing that tolerance is the hole: change “50%” to “56%” in a real Tesla 10-Q sentence and it is still 99.3% similar. In a round over five real 10-Qs, all ten doctored quotes scored between 99.1% and 99.7%, so a 95% match would have passed every one. Vouch keeps the tolerance for extraction noise but accepts a near-match only when every figure in the quote, and its scale, appears exactly in the matching passage. That rule slashed all ten.

the check · tesla 10-q
One changed figure, 99.3% similar
the claim's quote
Services and other revenue increased $1.54 billion, or 56%, in the three months ended June 30, 2026 as compared to the three months ended June 30, 2025.
similarity
99.3%
fuzzy match at 95%
passes
vouch
slashed: a figure is not the filing’s
Pick an edit or type your own claim. Similarity is difflib's ratio against the best aligned passage of the filing, the measure a fuzzy citation checker uses; at the usual 95% floor it lets every single-figure edit through. The figures rule only accepts a near-match when every figure in the quote, with its scale, appears in that passage.

Check the claim, not only the quote

Our own language-model analyst (Qwen 2.5 7B) quoted real sentences word for word ten times out of ten, and about half of its one-line summaries were still wrong: a six-month figure called a quarter, “Q3” for a quarter that ended in June, a gain described as a loss. A reader reads the summary. So the check covers the claim as written: the quote must be a whole sentence of the filing, every figure in the summary must be in it, the period must be that figure's own, and the direction may not be reversed.

The verdict and the money in one transaction

The contract accepts verdicts only from the Chainlink forwarder, and a slash happens in the same transaction as the verdict, so there is no window where a claim is known to be wrong and still holds its stake. The TypeScript check inside the workflow matches the Python one on 3,249 claims.

What ran

On Robinhood Chain testnet the market is deployed and its 50 claims are posted. The CRE workflow, in the CRE CLI simulation, wrote all 50 verdicts, 18 cleared and 32 slashed. The same claims settled end to end on a local devnet through the workflow's own code in Chainlink's SDK test runtime. The AI's wrong quarter is Qwen 2.5 7B's words from 24 September, replayed word for word. There are 57 contract tests, 76 Python tests and 24 for the workflow.

It does not prove that a view was honestly held, that a recommendation had a reasonable basis, that the filing itself is true, or that a record meets Rule 17a-4. That stays the firm's job; Vouch gives it evidence. It also cannot tell a borrowing limit from a balance, or an inflow from an outflow described in words, which is what the challenge window and human challengers are for.

Submitted to the BLI Legal Tech Hackathon on DoraHacks in October 2026, from an earlier version built for Casper and Arbitrum. The feed shows every claim with its filing passage and verdict; the code, the demo video and the entry are public.

Testnet only, with test USDG. The figure runs a port of the verifier's quote check in your browser on the Tesla 10-Q paragraph its tests use (accession 0001628280-26-049270); the full claim check, periods and directions included, lives in the repository.