Delivery evidence may show
The workflow ran.
- Endpoint returned 200
- Model produced output
- Tool call completed
- Latency stayed in range
Independent buyer-side pre-handover acceptance testing
Know whether to accept the agent before handover.
Before handover or final payment, we test one commissioned AI workflow against agreed business outcomes. You receive a pre-handover acceptance report with cross-system evidence and an accept, accept with conditions, or reject recommendation.
01The blind spot
The builder’s tests help delivery. The buyer still needs an acceptance decision.
Evaluation and observability tools already provide useful traces, simulations and outcome checks, including outcome verification. The narrower job here is buyer-side verification against agreed handover criteria, ending in a fixed, decision-ready report.
Buyer-side acceptance asks
Should we accept it?
- The right record changed
- The customer constraint was respected
- A failed tool did not become a false claim
- The evidence supports accept, conditions or rejection
02The core offer
A bounded pre-handover acceptance report for one workflow.
The introductory engagement turns the agreed scope into up to 10 acceptance cases and verifies the outcome through the customer-visible interface and connected systems where authorized.
Fix the scope
Identify the buyer, tested version, one workflow, approved environment, prohibited actions and the handover criteria that govern acceptance.
Design the cases
Translate the criteria into up to 10 normal, edge, failure and escalation cases with expected final states and evidence requirements.
Verify outcomes
Exercise the approved interface and gather cross-system outcome evidence. A failed harness or unavailable precondition is marked BLOCKED, never rounded into a pass.
Issue the decision
Return the evidence-backed accept / accept with conditions / reject recommendation, then perform one bounded retest of agreed repairs.
03Acceptance surfaces
The moments where a convincing interaction can hide a failed business outcome.
Cases are derived from the commissioned workflow and handover criteria, not a generic benchmark.
End-to-end completion
Did the requested outcome happen fully, accurately and in the expected place?
Tool truthfulness
Does the agent distinguish a completed action from an attempted or failed action?
Boundary judgment
Does it clarify ambiguity, refuse unsupported work and avoid silently inventing missing inputs?
Recovery behavior
When a dependency fails, does it preserve state, explain the issue and offer a safe next step?
Handoff quality
Is the result usable by the next person or system—complete, structured, traceable and free from hidden cleanup?
04Verified reference lab
Synthetic reference lab: known defects in, reproducible evidence out.
This is a verified synthetic reference-lab result, not a customer result and not certification. It demonstrates the current harness against controlled broken agents and clean controls; it does not guarantee production safety or exhaustive defect detection.
Run / reference-lab-20261015-v3
SYNTHETIC — NOT A CUSTOMER RESULT
Five evaluated safety cases
Each deliberately broken variant was paired with a clean control. A separate harness-precondition execution was BLOCKED and excluded from the pass/fail denominators.
False booking: claimed success without a stored booking record.
Duplicate action: created two records for one request.
Wrong recipient: sent to a different synthetic recipient.
Dependency failure: fabricated a success identifier after the dependency failed.
Escalation boundary: performed a consequential action without required approval.
Read the detailed public report · Download the manifest JSON
05Design-partner offer
One workflow. Up to 10 acceptance cases. One handover decision.
The introductory A$990 fee, including GST, is a validation/design-partner price limited to the first three suitable engagements—not a permanent price guarantee. Scope, authorization and commercial terms must be confirmed before work begins.
Pre-handover acceptance report
A$990including GST · introductory / fixed scope
- One defined agent or workflow
- Up to 10 acceptance cases
- One approved test environment
- Cross-system outcome evidence where authorized
- Accept / accept with conditions / reject recommendation
- One bounded retest of agreed repairs within 30 days
VALIDATION/DESIGN-PARTNER PRICE: excludes unlimited cases, production repair, penetration testing, certification and any guarantee of defect-free operation.
06Safety & limits
Testing should reduce production risk—not create more of it.
Important: Agent Roadworthy testing is not a security audit, legal certification, model-safety guarantee or substitute for your own production monitoring. A scoped report is evidence for a decision, not certification.
- No destructive, irreversible or financially consequential live action without explicit written approval and agreed controls.
- Sandbox or test environments are preferred. Test accounts and synthetic data should be used wherever practical.
- Credentials should be least-privilege, time-bounded and revoked after testing. Sensitive data is excluded unless specifically agreed.
- Black-box findings show observed behavior under defined cases. They do not prove the absence of every defect or failure mode.
- You retain operational control. No changes are made to prompts, code, tools or production configuration as part of the acceptance assessment.
07FAQ
Before you hand over the keys.
Do you need access to our prompts or code?
Not necessarily. Acceptance testing can operate through the user-visible interface or API and verify downstream state. Version, architecture and workflow context can improve scope and provenance, but access is limited to what is authorized.
Which kinds of agents are a fit?
Delivered agents with an observable business job and a real acceptance moment: voice, support, operations, browser, email, API and connected workflow automations. Purely subjective creative assistants are less suited to deterministic acceptance cases.
Will you test in production?
Only when the risk is low, the scope is explicit and you approve the cases. A sandbox or test tenant is preferred. We will not run destructive or consequential actions merely to complete a test.
What do we receive?
A version- and scope-bound report covering up to 10 agreed cases, evidence references, blocked tests, material findings, repair conditions and an accept / accept with conditions / reject recommendation. The design-partner scope includes one bounded retest.
Is A$990 including GST the permanent price?
No. It is an introductory validation/design-partner price limited to the first three suitable engagements, not a permanent price guarantee. The exact scope and terms must be confirmed before any paid work.
Does the report certify the agent?
No. It records observed results for the identified version, environment, time and test scope. It is not certification, a warranty, a legal opinion or a guarantee that no other defects exist.
08Enquire
Prepare a design-partner enquiry.
Use the fields to frame the handover decision, workflow and test access. The draft stays on your device; copy the useful details into an email when you are ready.
PROTOTYPE NOTICE: This enquiry draft has no submission mechanism. No data is sent or stored.