agentclaw

evaluation and governance

Governance that can show why a release is safe enough

Evaluation and governance connect policy to behavior. We define the operating boundary, run repeatable cases, record failures, and give the AI owner evidence for the release, revision, or stop decision.

evaluation and governance

Governance needs release evidence, not another policy file

A policy says what should happen. Evaluation shows what the workflow actually does when the input is incomplete, hostile, unusual, or changed by a new model. The CAIO needs both.

We turn expected behavior, prohibited behavior, and human review boundaries into repeatable cases. Each run produces evidence for a release, revision, or stop decision.

  • Success and failure definitions owned by the operating team
  • Cases drawn from approved workflows and known edge conditions
  • Regression and adversarial checks before a material change
  • A decision record with failures attached, not a score without context
sample · governance evidence

$ sample eval-suite --release-gate

scope: expected · prohibited · human-review

result: failures attached to source cases

owner review: required before release

decision: release, revise, or stop

what the evidence covers

Six questions governance cannot answer from a demo

grounding

Does it stay inside the approved source?

Test the missing-fact and conflicting-source cases where a plausible answer would still be wrong.

regression

Did the last change break an old promise?

Run the same cases after a prompt, model, tool, or policy change and attach the difference to the release decision.

tool use

Does it take only the permitted action?

Check arguments, permissions, skipped calls, duplicate calls, and the conditions that require a person.

adversarial

What happens when the input fights the rules?

Exercise instruction conflicts, hidden requests, data-exposure attempts, and actions the workflow must refuse.

edge coverage

Which real conditions have no tested behavior?

Record empty, malformed, multilingual, and exception-heavy inputs instead of treating the happy path as coverage.

measurement

Is performance moving in the wrong direction?

Keep the same operating cases over time so the monthly scorecard shows the change and the evidence behind it.

how evaluation runs

Define the boundary. Run the cases. Record the decision.

  1. 01

    Write the expected and prohibited behavior

    The operating owner and CAIO agree on what passing means, what must be refused, and when a person takes over.

  2. 02

    Build the repeatable case set

    Cases cover the ordinary workflow, known failures, boundary conditions, and adversarial inputs relevant to the real use.

  3. 03

    Attach the evidence to governance

    Every material change returns a result, failing cases, and an owner decision. The scorecard can point to evidence instead of a feeling.

Before evaluation becomes a release gate

Do you only evaluate systems AgentClaw built?+

No. The system can come from a vendor, another builder, or your existing stack. We need a permitted way to run it and enough approved examples to define the operating boundary.

Is a single score enough?+

No. A number without the case set hides what failed and why it matters. We keep the failing cases, the affected behavior, and the owner decision beside the summary.

How is evaluation priced?+

A one-off evaluation setup can fit a $1,500 to $2,500 starter build or a $5,000 production sprint, both fixed, depending on scope. Ongoing evaluation can sit inside CAIO Core or CAIO + Delivery after the ownership decision.

What access is required?+

Only the access needed to run the agreed cases and inspect the result. Sensitive examples stay subject to the company's approved data handling, and the decision log records the boundary.

Put evidence behind the AI release decision

The free AI ownership assessment starts with six questions. CAIO-qualified companies can then use an ownership-fit working session to decide where evaluation and governance belong.

It is not a free full roadmap or week-long audit.