Synthetic delivery sample

A release gate, end to end.

A fictional expense-reimbursement agent, twelve versioned cases, three recorded output snapshots and one reproducible release decision. Every result can be regenerated from the public fixtures.

Fictional workflow and data. This demonstrates delivery structure and evaluator behaviour. It is not a client outcome, testimonial or production benchmark.

Release comparison

Quality alone did not decide the release.

Candidate v2 removed the baseline's two critical failures, but its tail latency crossed the agreed threshold. It was held for review rather than waved through on a better pass rate.

VersionCases passedCritical failuresP95 latencyAverage costDecision
Baseline v18 / 1221,100 ms$0.01325BLOCK
Candidate v211 / 1205,200 ms$0.01250REVIEW
Candidate v312 / 1202,050 ms$0.01100PASS
Case coverage

Twelve cases built from observable policy boundaries.

The evaluator does not ask another model whether an answer feels safe. It checks decisions, required tool calls, argument values, policy citations, forbidden actions, latency and cost.

Missing receipt above the policy threshold
Manager approval required before reimbursement
Finance review for high-value claims
Forbidden expense category
Duplicate payout and idempotent retry
Missing required current-policy citation
Employee attempting self-approval
Negative and zero amounts
Inspectable delivery

Nothing important is hidden behind the report.

Open the policy, fixtures, evaluator, raw outputs and machine results. The PDF is the summary; these files are the evidence.

Evidence boundary

What this sample proves, and what it does not.

It demonstrates

  • A concrete case and output contract
  • Deterministic decision and tool-call checks
  • Quality, latency and cost gates together
  • A reproducible pass, review or block decision
  • A handoff another engineer can extend

It does not claim

  • A previous consulting-client result
  • Production traffic or customer data
  • A universal policy or evaluation framework
  • Model quality beyond these fixtures
  • Guaranteed outcomes for another workflow
Bring one real failure

We can test whether your workflow fits the same boundary.

Discuss one workflow