Verification infrastructure for AI agents

Agents propose. Gavel proves.

GenXis Gavel is an independent verification authority for AI-generated work. It runs the checks you declared, refuses to certify anything it cannot decide, and signs a receipt for every verdict, so "done" from an agent means something.

Never regress

State advances only when every declared check passes and nothing that passed before fails.

No self-certification

The model that proposes can never grade itself. Independent primary and shadow verifiers agree, or the answer is DENY.

A receipt for every verdict

Ed25519-signed, hash-chained, and replayable from the artifact root by anyone, offline.

For developers

Before your agent says it's done, Gavel checks the work.

Point Gavel at an artifact (a staged diff, a file, a test run, a JSON payload) plus a pre-committed oracle. It returns PROMOTE or DENY, the exact evidence, and a receipt you can verify without us. Call it from the CLI, an MCP tool, an OpenAI-compatible endpoint, or a pre-commit and CI hook. The same authority runs our public open-model benchmark at mostannoying.ai: 100 frozen tasks in five packs, every certified catch linked to its evidence.

Request

gavel.gate({ artifact: "staged-diff", oracle: "pytest", never_regress: true })

Gavel verdict

  • VerdictPROMOTE
  • Regressions0
  • ReceiptSIGNED

Hooks, CLI, MCP, API

Pre-edit, final-diff, and CI boundaries inside the agent workflow you already use, in observe, ask, or deny mode.

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 verdict re-runs from its recorded roots. Denials, duplicates, timeouts, and regressions are never billed as catches.

Join the developer waitlist

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.

01

AUTHORS NEVER JUDGE

Proposers propose. Only Gavel promotes. A model cannot score, certify, or bill its own output.

02

REFUSE OVER GUESS

Missing oracle, drifted source, mismatched hash, truncated response: Gavel stops and says exactly why. Unverified is never relabelled as pass.

03

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.

04

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.

These aren't slogans. They're gates the system fails closed on.

Company

Software that improves only when it can prove it did.

GenXis builds the control plane that lets AI systems get better without getting away from you: verified memory that keeps only what passed, counterexamples that compound into faster search, typed source-bound edits with machine-generated diffs, and improvement that has to beat its own control arms before it is promoted.

VERIFY

Independent primary and shadow checks against a frozen contract.

PROMOTE

Only a strict PROMOTE advances state. Everything else is recorded and denied.

COMPOUND

Verified lessons feed the next run. Control arms prove the memory is doing the work.

Latest controlled run: 12 of 12 unseen tasks with verified memory, 3 to 4 of 12 without it. Zero retries, $0.04 spent, every result under a Gavel receipt.

Build with us

Every verdict signed. Every improvement proven.