Clavenar
Champion track · 2 of 3

Run a four‑week evaluation.

Evaluations that try to cover every agent and every tool fail because they take months and convince nobody. The Founding Design Partner Offer is exact: four weeks, one agent, one tool/action surface, and a dated decision and exit.

Scope

  • One agent. Pick the loudest one — the one your team already worries about. Don’t pilot on the safe agent.
  • One tool surface. A single class of calls (e.g. “send email,” “create ticket,” “wire transfer”). Twenty tools is post‑pilot.
  • Approval flow rehearsed. Exercise Yellow‑tier human approval with test fixtures before any separately approved production step.
  • Observe before narrow enforcement. Production and enforcement require explicit written customer approval in the signed Order Form; the offer alone grants neither.

Success metrics

  • Verdict latency p95. Set the acceptance target after an observe‑mode baseline; measure cache hits, misses, and live semantic-provider calls separately.
  • Blocked attacks. Run the chaos catalog at the pilot agent; count the verdicts.
  • HIL median approval time. A signal that the approval flow is sustainable, not a bottleneck.
  • Ledger chain verification. Tampering a row should break /verify. The chain should pass without it.
Week by week

What the four weeks look like.

  1. W1

    Scope and baseline

    Confirm one designated agent, one tool/action surface, owners, data classes, success criteria, and exit evidence. Establish an evaluation baseline with approved test fixtures before changing any real workflow.

  2. W2

    Observe

    Run the agreed scope in observe mode. Review what would be allowed, parked, or denied; separate cache and live inspection latency; and tune noisy policy without authorizing side effects.

  3. W3

    Tune and rehearse

    Run the chaos catalog and rehearse Yellow‑tier approvals with test fixtures. If the signed Order Form separately authorizes a narrow production slice, activate only that written scope.

  4. W4

    Decide and exit

    Review the agreed metrics, revoke evaluation access, and retain the promised deletion or return evidence. Stop, extend by signed amendment, or execute a new production Order Form; silence does not convert the evaluation.

What you’ll be able to decide in week 4

The outcome story.

“For the one agreed agent and tool surface, we can show observed verdicts, approval rehearsals, latency by path, and chain verification against the success criteria we signed before the evaluation.”

Previous ← Why your team needs guardrails
Next — 3 of 3 From evaluation to production →