REAL REHEARSAL · HONEST CHECKPOINTS

See what was actually tested.

These are real screenshots of the generated app, not design mockups. Captions state exactly which checkpoint each image proves.

Real Lovable-generated OpsDesk app showing synthetic tickets, search, priority filter, create form and an explicitly disconnected AI panel — Real browser capture · 31 October 2026. The generated app passed create/search/filter/select interactions and displayed an honest disconnected-AI state. This image is the first-build checkpoint; it does not prove a live AI call or trace ingestion. Pixels reviewed: synthetic content only; no account name/email, credential or private URL appears.

END-TO-END ACCEPTANCE IS A SEPARATE CHECK

A full run must create/select a ticket, obtain an actual native-model/tool answer, demonstrate the server-side change stop and independently find its exact trace in the configured Langfuse project. Local tests, generated source and this screenshot cannot substitute for that proof.

Current rehearsal evidence

Three learner paths and 18 persona-specific part workflows/Markdown briefs are authored.
Reviewed local contract rehearsals pass for all three persona routes; these use explicit model doubles/source checks.
The Lovable-generated UI was tested in the actual service using synthetic tickets.
Approved app sign-in completed. The actual signed-in preview returned an answer with two native model calls, service_health and release_diff tool events, and 952 reported tokens. The change-request control returned approval_required, zero model calls and an explicit server-block event. These are observed app responses, not a complete source security audit. The downloaded app source was independently tested locally: all 31 tests passed, including eight exporter tests and eight server-safety tests; these use isolated doubles, not a live telemetry backend.
One fresh native Lovable run was independently matched in the intended EU Langfuse project: exact trace and random session, six observations, two model calls, two synthetic tools and 948 reported tokens. All six observations had no prompt/answer input or output.

The native-AI and guardrail screenshots below are real captures with reviewed pixels. The reviewed trace tree below records independent provider readback. Your own app and credentials still require your own acceptance checks.

Choose your route → · Connect your Langfuse instance →

From a ticket to a real answer

Actual authenticated OpsDesk answer explaining synthetic latency, uncertainty and safe next checks — Actual Lovable preview · native AI completed in two model calls. The expanded inspector separately reported both tools and 952 tokens. This selected capture excludes gateway identifiers, account details and authenticated preview URLs. Langfuse is explicitly not configured.

Stop before a change

Actual approval_required result with zero model calls and an explicit server-side block event — The same authenticated app's change-request control returned approval_required, model not called, zero calls and the server-block event. This check covers this explicit change path; it does not establish universal prompt-injection protection.

Actual agent trace in Langfuse

Actual matched native run: agent, authorization, two generations and two tools — Independent provider readback · six observations · 948 reported tokens. Account controls and private identifiers excluded. Displayed costs are provider calculations, not billing receipts. This does not prove OCI APM or Log Analytics delivery.