AmabieGitHub Get started
AI AGENT VERIFICATION LAYER

Your agent says done.
Amabie checks the change.

Independent verification with your project’s own checks and a sealed baseline. The verdict comes from evidence.

See how it works Read the docs
Independent
verification Deterministic
verdicts Inspectable
evidence
AmabieVerification runRUN 001
ILLUSTRATIVE DEMO Browser only
  1. BaselineSealed starting point
  2. CandidateProposed change
  3. ChecksDeclared project checks
  4. EvidenceCompare observations
  5. VerdictDeterministic result
baseline01
DEMO / base
12 passed
0 failed
candidate02
DEMO / change
12 passed
0 failed
DEMO VERDICT
NOT PROVEN

Green on both sides. These checks do not distinguish the change from its baseline.

Check observationsILLUSTRATIVE
MEASURED SIGNAL

Project checks observed

No discriminating check

Evidence available below

BEFORE → AFTERbaseline.failed0candidate.failed0

Task: Change behavior without a check that measures it

No code is executed or uploaded. Sample counts explain the product’s verification model.

Demo verdict: NOT PROVEN
BUILT AROUND THE CHECKS YOU ALREADY OWN
Node / TypeScriptPythonRustGoCommand-driven projects
01 / THE VERIFICATION MODEL

From claim to verified evidence.

Run the project’s checks outside the agent’s conversation. Compare the change with its sealed starting point.

1

Seal the baseline

Pin the starting state, verification plan and tool identities.

2

Run the checks

Measure baseline and candidate with the declared checks.

3

Compare evidence

Inspect differences. Reproduce divergences.

4

Issue a verdict

Retain the result and the evidence behind it.

01 / THE QUESTION BEHIND THE GREEN CHECKCHANGE ≠ PROOF

Why green tests are not enough.

Would these checks pass without the change? If yes, a green result alone does not establish evidence for the new behavior.

BASELINE

The code before.

All declared checks pass
KNOWN STARTING STATE
CANDIDATE

The agent’s change.

The same checks pass
NEW CODE, SAME SIGNAL
AMABIE ASKS

What did the
change prove?

NOT PROVEN

Checks that cannot distinguish a behavior change leave a proof gap.

A failure-sensitive check can establish evidence. A binding declares coverage. Neither makes Amabie understand every requirement.
02 / A DOCUMENTED PROOF CASE HISTORICAL FIXTURE

A real regression. In the evidence.

Axios 0.27.2 → 1.0.0. Same downstream suite: axios-mock-adapter, pinned at b8804442.

AXIOS / 0.27.2 → 1.0.0REPRODUCIBLE EVIDENCE
BASELINE · AXIOS 0.27.2
128passing tests
0 failing2 baseline rounds
CANDIDATE · AXIOS 1.0.0
125passing + 3 failing
3 failing3 candidate rounds · unanimous
CONFIRMED REGRESSION

Headers matching and baseURL resolution broke.
Dependency-tree drift was confined to the Axios subtree.

Inspect the documented case

Documented results, not a live run on this website. Host-exact proof assertions require the committed proof host; other runtimes report incomplete coverage explicitly.

03 / INSIDE THE VERIFICATION LAYEREVIDENCE FIRST. EXPLANATION SECOND.

The model writes. Evidence decides.

Your tools execute the checks. A deterministic classifier issues the verdict. Explanation sits outside that decision.

01

Seal the starting point

Pin the baseline, declared verification plan and tool identities.

BASELINE + PLAN
02

Run and compare

Execute checks against both states. Compare observations and reproduce divergences.

CHECKS + OBSERVATIONS
03

Classify the evidence

Derive the verdict and retain a machine-readable Evidence Bundle.

VERDICT + EVIDENCE
Optional AI adapter · shipped provider is a no-op, not a live AI service.
HONEST OUTCOMES

A result can be uncertain.
A skip is never a pass.

PASS+

The measured evidence supports the change.

CONFIRMED REGRESSION+

A baseline behavior fails in the candidate and is reproduced.

PRE-EXISTING FAILURE+

The failure is already present in the baseline.

FLAKY+

Repeated outcomes do not agree.

INFRASTRUCTURE FAILURE+

The execution environment failed.

INCONCLUSIVE+

The evidence cannot support a stronger result.

04 / FITS THE WAY YOU WORKWorkflow guide

Own your workflow.
Inspect the finish line.

Automatic completion checks for everyday work. A separate candidate when isolation matters. The same clear account of what was actually measured.

01

Your everyday workflow

Discover the project’s declared checks, seal the plan, and wire completion hooks. The agent registers each new request through MCP; Amabie checks the result when the turn ends.

amabie setup --yesIMPLEMENTED
02

An isolated candidate

Work in a separate candidate when changes need isolation. Objective requirements need sealed check bindings. A subjective duty stays with a human reviewer.

amabie work → amabie finishEXPERT WORKFLOW
03

Inspect the evidence

Read the current state without running checks. Use doctor for a fresh check, or the proof pipeline for an attested baseline/candidate comparison.

amabie result --jsonIMPLEMENTED
INTEGRATIONS, WITHOUT THE WISHFUL THINKING

Capabilities,
clearly labelled.

The hook, the trust step and the actual limits all matter.

Claude CodeGATED

Completion and pre-tool hooks. A new request must be registered; unbound criteria remain not proven.

OpenAI Codex CLITRUST STEP REQUIRED

Completion hook installed by setup. Review and trust it once in Codex before it can gate completion.

Other command-line agentsADVISORY

Machine-readable status and checks. Reporting does not mean Amabie can block the agent.

05 / TRUST, WITH THE BOUNDARIES INCLUDEDCURRENT LEVEL: LOCAL

A clear verdict.
An honest boundary.

Verification is only useful when you can inspect what stands behind it.

Read the trust model
THE AMABIE WITNESS

Watch the change.
Keep the evidence.

Local authority, sealed outside the project

Ed25519-signed records are kept outside project trees. The same user identity can still replace the store, keys or history, or reseal a different plan.

Evidence coverage is not semantic truth

A bound check measures a declared obligation inside Amabie’s model. It does not guarantee that the check fully expresses the requirement.

HARDENED is not available in this build

Provider and broker code exists. A separately measured authority boundary is required before a hardened level can be claimed. It is not activated here.

IMPLEMENTED PROVIDER · UNAVAILABLE CAPABILITY

Unknown evidence stays uncertain

Unobserved runners have a weaker evidence ceiling. Amabie does not promise universal fail-open prevention, semantic correctness, or token savings.

THE NEXT AGENT TURN DESERVES EVIDENCE

Make “done”
mean something.

Read the setup guide Review the Amabie 2.0 candidate

The Amabie 2.0 package is still in release validation. Public releases may retain the Canary name until the rename is released.