Release rehearsal
Paste a staging URL and tell three NVIDIA Nemotron visitors what a first-time user should accomplish. Watch each phone-sized visit, then keep the screen and path that explain where the release got stuck.
Pick a recorded run.
Recorded playback · each colour is one agent · squares are steps · lines are paths · rings mark a stuck ending
| Visitor | Said done | Screen agreed | Job done |
|---|
Evidence library
The comparison set contains 122 real hackathon apps. It separates an app defect from a model getting lost and from a limit in this simulator, so a stuck visitor is not automatically blamed on the product.
Observed collection: 122 apps, five failure types, three reproduced fixes. Visitors are models, not people.
These are command-line agents on controlled historical CPython fixes, separate from the browser swarms. Ten matched cached runs on each side use the same frozen checker. Solid dots passed. Hollow dots failed. The guide case below had no improvement.
Browser swarms, command-line cases, a first-visit study, and earlier product checks. Each answers a different question.
64 met the clean GET gate. 43 were untested or skipped by policy. 93 were not green; that is not a count of broken apps.
One saved read-only visit per app on 16 October. No sign-up, form or payment task. Human labels: 0 of 30. Timings are local one-shot observations.
Read the full method and top 20 · Ten descriptive patterns
Recorded on 17 October. A basic browser check plus one Nemotron Nano explanation through Nebius Token Factory. The saved report survived a service restart.
This explains checks that already ran. It does not show a new agent completing the task. Model completion status and raw stop reason are preserved in the local integration.
All five agents reached the working guide, quoted its navigation label, then got lost. The harness silently dropped short quotes. It now gives explicit rejection feedback and separates page content from controls.
The repaired rerun still completed 0 of 5. The phone run broke off after a provider error. Super VL rejected a guide reading claim; the evaluation limitation remains open. No target-app bug was proved.
Read the trajectory audit13 October, before this studio and the five-agent runs. Historical local checks and replay, not a fresh swarm or proof of forecast accuracy. The previous loop reports and screenshots remain in the project inventory.