Synthetic demo
Synthetic fixtures let you learn and certify the RunOnProof decision contract without making claims about a real company. They are fixed examples with known expected behavior.
Why synthetic comes first
Real company data changes, sources can be unavailable, and business consequences can be high. A synthetic case isolates the integration: if the agent mishandles REVIEW or duplicates a request, you can fix the workflow before real onboarding or payment is involved.
Seven essential fixtures
| Fixture | What it proves |
|---|---|
| Active company | The happy path preserves entity, evidence, scope, and decision |
| Ambiguous company | The agent requests a stronger identifier instead of guessing |
| Source unavailable | Missing evidence does not become approval |
| Conditioned capability | A documented dependency remains visible |
| Payment required | Commercial handling is separate from business approval |
| Idempotency | Repeating the same intent does not create a second execution |
| Replay | A committed result can be returned without pretending it is fresh |
Expected agent behavior
For each fixture, record the tool input, top-level decision, explanation shown to a person, downstream action, decision identifier, and whether an economic event occurred. The test passes only when both the response and the agent's behavior are correct.
Example: ambiguous company
Input contains a common company name with no registration number and two plausible matches. Expected behavior: REVIEW. The explanation asks for a company number or stronger identifying information. The agent does not select the first search result.
Example: payment required
The commercial layer indicates that payment or approval is required before execution. Expected behavior: the agent pauses, shows the price and required action through the approved flow, and does not label the company PROCEED or STOP based on the commercial response.
Example: replay
The same completed action is requested with the same idempotency key. Expected behavior: return the committed decision with its original observation time and no second charge. Replay is not a fresh source read.
Certification record
Keep the fixture version, integration version, RunOnProof contract version, test time, actual result, expected result, and reviewer. Rerun certification after relevant contract, policy, state-mapping, or agent-prompt changes.
Moving to real data
After all fixtures pass, test one controlled workflow where a human still owns the final action. Compare the legal entity, evidence, limitations, and decision behavior. Do not jump directly from a happy-path fixture to autonomous payment.