Sondel product lab: saved results walkthrough

This transcript describes a synthetic local lab. It is not a customer result or a live widget recording.

Scene 1
This is a walkthrough of saved API results from the Sondel product lab. It used synthetic Northstar data. It is not a customer result or a live widget video.

Scene 2
The lab used the local Sondel application, an isolated PostgreSQL database, Voyage retrieval and reranking, Baseten DeepSeek drafting, and StrictGate checks. It bypassed the browser widget, owner dashboard, visitor authentication, production services, action authority, and external delivery.

Scene 3
The first paid run saved six cases, then failed during a database content-hash query. Its cancellation question received a safe handoff instead of the expected cited answer. The record also preserves a false automatic check that manual review corrected.

Scene 4
The completed run changed the fixture and question. It asked about reopening a resolved support request, not cancellation after purchase. The fourteen-day answer was correct for that submitted question, but it did not test the original cancellation intent.

Scene 5
Other saved outputs kept useful limits visible. The model did not invent a nonprofit discount or a backup-storage country. It declined an unauthorised cancellation request, stored needs person, and offered handoff. No action or handoff was executed.

Scene 6
After the synthetic reopening policy changed, the saved answer changed from fourteen days to thirty days, with a citation. This demonstrates the reopening source update only. It does not complete the original cancellation test.

Scene 7
The two paid attempts kept thirteen outputs. Their combined known application estimate was zero point zero one nine two six three US dollars. The full JSON preserves every prompt, output, source, review, timing, cost, scope change, and failure. There is no aggregate original-plan pass claim.

Full record
The complete saved API corpus is available at /assets/product-lab-2026-10-23.json.

Bypasses
The recording and lab bypassed the browser widget, owner dashboard, visitor authentication, production services, action authority, external action hosts, handoff submission, and email delivery.