Browser fixture adapter, strict result validation, source checks, human review, abort/timeout tests and a lazy Vue annotation reader.
02 / AI integration workbench
Evidence before
automation.
Turn a service ticket into a reviewable next step. Inspect the sources, test the contract, then decide what happens.
Ticket → grounded draft → human → adapter
A small workflow. Explicit boundaries.
01 / Input
Synthetic service ticket
12–2,000 characters. Document instructions are untrusted data. Text and drafts stay in this tab’s memory.
Ready. Four fixture attempts per ten seconds.
02 / Contract
A grounded draft
Start with a ticket.
The adapter returns a typed classification, recommendation and exact source spans. Invalid output never reaches review.
Inspect the output schema
03 / Provenance
Read the source material
Locally authored synthetic documents, revision 1. Citation links highlight exact spans. Matching text does not establish a recommendation’s correctness.
04 / Human decision
You decide. The adapter waits.
No customer contact or real service API is connected.
Generate a valid, actionable draft to begin review.
Open this view to load the local Vue runtime.
Measure the contract
Failures belong in the test set.
Reproducible synthetic cases measure fixture behavior, not model quality or production throughput.
No evaluation run in this tab yet.
| Case | Result | Elapsed (ms) |
|---|---|---|
| Run the evaluation set to populate this table. | ||
Method: performance.now() around each assertion, sequentially in this browser. Cancellation and timeout tests use accelerated 30/10 ms fixtures; service-desk latency is intentionally 650 ms. Timing depends on this device and is not a service-level claim.
Engineering evidence
Prepared, testable, honest.
The optional server-side Responses API adapter. It needs approved infrastructure, authentication, budget and review before enabling. No key belongs in this page.
Live model reliability, people leadership, paying client references, high-traffic operations or RFP compliance.