KaapDocsCheck a treasury
KaapDocs
What to trust

Limits

What Kaap does not do, does not catch, and has not built yet.

A guard you cannot see the edges of is not worth trusting. These are Kaap's.

What is not live

  • Mainnet. The guard and the registry are on NEAR testnet only. Kaap reads mainnet treasuries, and that is all it does there.
  • The always-on watcher. It is hosted on Railway since 13 Nov 2026 for this deployment; on this host since 12 Nov 2026 and watches the demo treasury. Its status page is public. No other treasury is enrolled, and it runs on free data providers, so a vote can be held when one of them does not answer.
  • Strict mode on testnet. The guard has strict mode and the 24-hour delay on key changes since 13 Nov 2026. No treasury on testnet has turned strict mode on and no key change was announced, so both are proven by tests against the real Sputnik DAO binary, not by a live run.
  • One build of the reader at a time. The reader on testnet is the build of 13 Nov 2026, made from the same rules as the check on this site. When the rules change, the site's check runs them at once and the reader only after a new build is published, so in between the two can rate a few requests differently.

What Kaap does not catch

  • A theft that looks like an honest payment. If nothing in the request proves it is a theft, plain code cannot justify a veto. In the evaluation Kaap missed 13 of 54 planted attacks for this reason. See Evaluation.
  • A request it cannot read. A request that shows only a hash is flagged to members as unreadable. It is not vetoed, unless a member supplies the words behind the hash and they turn out dangerous.
  • A request that passes in one step. The veto has to land before enough approvers sign. In the recorded proof that took 24 seconds. A treasury whose rules let a single signature execute a request leaves no time for any guard.
  • A very large request. A rule change of more than about 200 KB is too large for the guard to vote on in one transaction.
  • A change that needs no request. Kaap guards requests. A change to a treasury's own code that does not go through a request is outside what any veto can stop. Kaap notices it while a request is open: it judges the open requests again, and stops watching a treasury whose code is no longer a Sputnik DAO release, which shows on the watcher's status page. kaap verify shows a code change since a verdict. Members get no alert for the code change itself.
  • The same request, filed again. A rejected request is closed for good, but nothing stops a member from filing it again. Kaap reads the new one afresh, and rejects it again if it is the same theft.

What you still have to trust

  • Kaap's founder, for the guard's key. The guard's code is locked, but its owner can replace the key it obeys. A change is announced on chain and waits 24 hours, the watcher tells members at once, and the owner can cancel it. After that day a dishonest owner could point the guard at a key of their own and reject requests. They still could not approve or move anything. The owner key is one key with no second signer.
  • OutLayer and Intel's sealed hardware. A verdict is only as good as the promise that the reader ran unmodified. It also needs OutLayer to answer: when its testnet is busy a sealed run can go unanswered, and the watcher asks again for as long as the request is open. On 13 Nov 2026 one request waited about nine minutes for its first verdict.
  • The watcher's key. The key on the watcher's host can run the sealed reader and pay for it, and nothing else. The reader is published under a separate owner account whose key is offline, and OutLayer refused all five of the watcher's attempts to publish a build, choose the active one, remove one, take the project or delete it (scripts/attack-sealed-owner.mjs, output in PROOF.md). Someone who stole the watcher's key could still spend its balance on real verdicts and on the vetoes the rules allow.
  • Two RPC providers. Kaap reads the chain through two and refuses to judge when they disagree. If both were wrong in the same way, it would be wrong too.

What has not been checked

  • No independent security audit has been done.
  • The evaluation is 373 requests, labelled by AI passes against a written guide. It is enough to see the shape of what Kaap catches, not to promise a rate.
  • A wrong reject costs a treasury time: the request has to be filed again. In the evaluation that happened to none of 244 honest requests, which is not the same as never.