VERIFY
Separate verifier checks evaluate the declared contract; unresolved evidence remains unverified.
Verification infrastructure for AI agents
Agentspropose.GavelTM proves.
GenXis GavelTM is an evidence-bound verification gate, separate from the proposing agent or system. Point it at a named artifact and declared check. It returns a bounded decision and a signed receipt that says exactly what was verified. Unverified is never a pass.
Mission
The war on slop is won at the gate.
Slop is work that looks finished and is not. GavelTM gives teams a clear boundary before they rely on an agent completion claim: the agent proposes, a separate gate evaluates the declared evidence, and people or protected CI retain the promotion decision. Missing or unresolved evidence stays unverified rather than becoming a green signal.
PROMOTE
A bounded decision is recorded. Your human or protected CI policy decides what advances.
COMPOUND
Retain receipts and findings so your team can inspect and address repeated bounded failures.
Never regress
State advances only when every declared check passes and nothing that passed before fails.
No self-certification
The proposing agent or system does not grade its own submission. A declared verifier evaluates the named evidence, or the result remains unverified.
A receipt for every verdict
Signed and bound to the named artifact and declared evidence. Offline verification validates the recorded signature and binding; it does not rerun the check, establish verifier diversity, or expand the original claim.
For developers
Before you rely on an agent's completion claim, check the named evidence.
Point GavelTM at a named artifact (such as a staged diff, file, test output, or JSON payload) and a declared check. It returns a bounded decision, the recorded evidence, and a signed receipt you can verify offline. It receives only the artifact you explicitly submit for remote verification; it does not crawl your project. Call it from the CLI, an MCP tool, an OpenAI-compatible endpoint, or a pre-commit and CI hook. mostannoying.ai is a public example of bounded, recorded evaluations.
Request
gavel.gate({
artifact: "staged-diff",
oracle: "pytest",
never_regress: true
})
Illustrative GavelTM verdict
- VerdictPROMOTE
- Regressions0
- ReceiptSIGNED
PROMOTE in this illustrative example covers the named artifact and declared check only. It does not establish complete application security, runtime correctness, or coverage outside the named evidence.
Hooks, CLI, MCP, API
Optional observe hooks and bounded checks in the agent workflow you already use. Protected CI and your review policy remain the enforcement boundary.
Decidable oracles
Commands, test suites, JSON schema, exact output. Every oracle is hashed and frozen before the run. No oracle, no verdict: it refuses rather than guesses.
Replayable receipts
Every receipt records its subject and evidence binding. Offline verification checks the receipt; it does not rerun the original check or prove anything outside it. Denials, duplicates, timeouts, and unverified outcomes are not billed as catches.
Our principles
Verification you can't talk your way past.
Every rule below is enforced in code and shows up in the receipt, not on a slide.
AUTHORS NEVER JUDGE
Proposers propose. GavelTM evaluates only the named evidence against the declared check. A model cannot certify or bill its own output.
REFUSE OVER GUESS
Missing oracle, drifted source, mismatched hash, truncated response: GavelTM stops and says exactly why. Unverified is never relabelled as pass.
EVIDENCE OVER ASSERTION
Raw provider bytes are stored before they are parsed. Every dispatch, retry, and cent lands in a hash-chained ledger you can audit.
BOUNDED BY DESIGN
Nothing runs without a signed owner authorization bound to one exact proposal hash. One shot, no silent retries, explicit spend and wall-clock caps.
Build with us