NANO DATA

Inspiration

When work moves from one AI session to another, people often have to explain it again. NANO DATA explores a practical alternative: keep the original record, preserve the purpose and reasoning, and pass a source-linked handoff so the next AI can continue the task.

What it does

The working seed connects an original record, a purpose, a reason, a next action, a CORE capsule and a stored receiver result. The original stays available for review rather than being replaced by a summary.

Our demonstration uses one fictional reading-club example. The receiver produced a preparation checklist that retained two tables and six chairs and kept attendance, satisfaction and outcomes unmeasured. The workflow screen displays the source, CORE and result, and reloads the saved result. Preparation suggestions are recommendations, not claims that those actions were completed.

How we built it

The prototype uses NVIDIA Nemotron 3 Nano 30B A3B served through Nebius. Producer and receiver are separate requests to the same configured model, not independent model providers. The selected producer and receiver receipts report HTTP 200 responses and saved outputs. The application links the records so a reviewer can inspect the original and the continuation together. Provider-reported receipt fields are not independent provider telemetry.

What the video demonstrates

The 150-second English-captioned video is a replay of a saved real-app trial, not new inference during recording. It shows the original, purpose/reason/next action, CORE, the receiver checklist and saved-result reload. The useful receiver result came from an additional receiver trial; this is not a first-attempt success claim. Review-required labels remain visible.

Challenges and lessons

A response can arrive successfully without preserving the meaning of its source. We therefore separate transport success, persistence and semantic review. The important question is whether the next AI follows the intended task without turning uncertainty into a false fact.

Scope and next steps

This is one synthetic handoff, not a general accuracy benchmark. It does not establish superiority, lossless compression, cost savings or a finished autonomous chain. The larger vision is a chain of source-linked decisions; the demonstrated contribution is its first inspectable step.

The public source is available at https://github.com/shoyuman-nano-data/nano-data as an MIT-licensed source archive with setup instructions. The published archive now includes the final-mode costguard code and matches the recorded code-only deployment; final judging remains disabled. Dedicated provider connectivity and final judging-period access are not yet ready. The video shows historical saved work, not a successful run of the final judging configuration. This draft does not claim that every submission requirement is complete.

Built With

• nebius
• nvidia