Inspiration
Teams often approve a policy in prose and discover its real effect only after code reaches production. A CCPA records export makes the gap concrete: if an analyst queues a request and their access is revoked before delivery, should the system honor the earlier authorization or check again when it releases the records?
What it does
Trajectory is a native iPhone decision workbench. The owner selects the governing authorization rule, and NVIDIA Nemotron 3.5 Lightning through Nebius Token Factory constructs a bounded Go candidate for that choice. The app shows the changed source and keeps the parent and previous candidate visible. The owner then starts a separate local check: Go build, vet, nine protected tests, and a request → revoke → deliver scenario. A construction is never described as a passing check before that check runs.
In the recorded run, request-time candidate A exported after revocation, while delivery-time candidate B-2 blocked release. Both passed the protected check suite. The candidate source, response IDs, digests, and check summary are published with the app repository.
How it works
The Mac companion imports a public synthetic service. It sends the decision and protected source context to the Nebius Token Factory chat-completions endpoint using nvidia/Nemotron-3_5-Lightning. Only queued/queue.go and queued/worker.go may change. The companion independently checks the full source boundary before exposing a candidate in the app. It stores the model response identity and source digests, then runs the checker against the exact candidate tree and accepted rule. The API key remains on the Mac host.
What was built during this event
BMO's aggregate engine and public decision-guided synthesis design predate the event. During the submission period, I built Trajectory's native workbench and standalone companion. On October 30 I added the direct Nebius Token Factory provider, ran real Nemotron A/B constructions in the native session, added model provenance to the UI, published the two generated candidates and independent check evidence, and produced an iOS Simulator test build. This standalone demo does not claim a live BMO runtime integration.
Challenges and learning
The first model response needed a larger output budget. One intermediate B construction did not return the required two-file shape and was rejected before publication. The successful B-2 candidate passed the immutable source boundary and independent checker. This is why Trajectory keeps generation, acceptance, verification, and commit as distinct actions.
Nebius feedback
Token Factory's OpenAI-compatible chat-completions API made a bounded provider integration straightforward. The returned model identity, response ID, and token usage helped bind the generated files to a real inference call. For coding workflows, structured-output reliability and transparent reasoning-token accounting matter: malformed file shape must fail closed, and a 5,500-token output budget was insufficient for one response even though the final code was small.
Log in or sign up for Devpost to join the conversation.