Example deliverable / Not a client assessment
An assessment you can inspect.
A fictional insurance-claims workflow was tested against a fictional policy oracle. The candidate was an intentionally incomplete deterministic rule-based simulator. No customer system or real LLM was tested.
Executive summary
We record a decision path, mismatches and proposed engineering remedies. All failures were deliberately seeded into a toy simulator; they are not customer incidents or evidence of real AI model performance.
Findings and proposed controls
| Finding | Fictional failure mode | Control to validate |
|---|---|---|
| F-01 | Stale evidence or policy version | Bind decisions to source hashes and effective policy versions |
| F-02 | Duplicate or uncertain tool actions | Reconcile execution state and idempotency keys before retry |
| F-03 | Untrusted documents acting as instructions | Keep document content outside tool authority |
Inspect the evidence
Each fictional case has a recorded expected outcome, toy output and comparison. The frozen fixture checksum is published alongside the 30-case table.
Explore the case-by-case evidence
What a real assessment would add
Under an agreed scope: actual agent outputs or supplied traces, replayable findings, confidence and severity labels, and a ranked fix list. The $1,000 introductory offer covers one agreed workflow; no sale or outcome is implied. Typical delivery is 3–5 business days after agreed materials arrive.
This example is not a penetration test, certification or safety guarantee.