A control plane that recommends and never deploys.
Your whole team reads the same conversation-QA state from a workspace that runs in your own environment and ships nothing anywhere.
$ hotato serveOne instrument, five reads on the same registry.
Every view has a machine mirror at ?format=json (same auth, same data) so an agent or a script can drive the workspace without scraping HTML.
Release readiness
Per-release rollup: whether required suites are complete, failures by dimension, the inconclusive count, the recorded-vs-simulated split, and new-vs-fixed since the last release.
Scenario matrix
Rows are scenarios, columns are this release and the last, with a per-dimension status and pass^k where a scenario has repetitions. Filter by agent, release, suite, and status.
Conversation inspector
One call's evidence manifest, transcript, trace spans, and per-dimension evaluations, deterministic and advisory shown in separate lanes. Every digest links straight to the raw evidence blob.
Failure clusters
Failed checks grouped by observable signature (dimension, assertion kind, reason-class) with drill-through into the inspector. It groups what was observed; it does not claim a cause.
Production health
Ingest coverage and per-dimension failure rate over time, computed separately for recorded and simulated calls. Thin history reads not enough history.
Runs in your environment
SQLite registry plus a content-addressed evidence store. Bearer-token auth, no per-agent cap, no phone-home.
Self-host the workspace →In your VPC, on loopback, phoning nothing home.
The server binds 127.0.0.1 unless you tell it otherwise, authenticates every request against one bearer token compared in constant time, and opens only a listening socket.
It is stdlib-only: http.server plus sqlite3, no framework, no build step, no telemetry. A zero-egress test whitelists loopback and fails if any view reaches for an outbound connection.
$ docker compose up -d # http://127.0.0.1:8321SELECTs. Reviews and labels stay CLI-driven; the only file it writes is an append-only audit log.Separate, flagged, never blended.
The workspace surfaces the shape of the evidence. Real and simulated calls are computed apart, thin samples are marked, and no view rolls the whole thing into one grade.
Redacted transcript segments render as [redacted] in both the HTML and the JSON mirror, and every request appends one line to an audit log that records a token prefix, never the secret.
- recorded vs simulated
- Kept separate everywhere. A simulated call is never merged into a bucket of recorded ones.
- small samples
- Flagged in place as
low sample, N=3, never smoothed into a confident-looking line. - combined quality number
- Absent by construction. No blended score, anywhere.
- write endpoints
- None in v1. The server mutates nothing but the audit log.