AIZOYA ProofOS

Inspiration

AI can produce a convincing answer without showing whether the evidence supports it. We built ProofOS to make that gap visible: separate what ran, what the evidence supports, and what still needs review.

What it does

ProofOS reviews a claim against operator-supplied source text. A deterministic passage check tests whether the quoted words occur in the supplied excerpt. An owner-local model assessment uses NVIDIA Nemotron through Nebius Token Factory to evaluate support for the claim. The application produces downloadable receipts with input hashes, checks, incomplete checks, model attribution, execution evidence, and a recommended next action.

The current inference workflow supports one fixed prepared sample. It is an evidence-review prototype, not an autonomous research or business-action system.

A real result: matching words were not enough

Our supplied passage said, “Submissions are due Friday, October 30 at 10:00 AM PT.” The proposed claim additionally asserted the year 2026. The quotation check passed, but Nemotron returned INSUFFICIENT because the supplied passage did not state the year. ProofOS retained NEEDS_REVIEW instead of declaring the entire claim verified.

This result came from our completed owner-local runtime request, not a fixture model response. The original source was not fetched or authenticated by the application.

How we built it

Python implements deterministic input, evidence, authorization, persistence, and receipt checks. JavaScript provides the interactive interface. NVIDIA’s nvidia/nemotron-3-super-120b-a12b is called through the Nebius Token Factory chat-completions API for the model assessment. SHA-256 binds source candidates and sample input to inspectable records.

Vercel hosts a publicly accessible interactive preview of the passage-check interface and the recorded original Nemotron result. The hosted source is commit ac0ee278588e89396536f0f538d6dc324e0dcc2e, release sha256:2f56934dd3ff1b8ef4c9e623aa65328122122fc280d4dffa94b708b22ff66eff. The completed inference ran on the owner’s local machine with a private server-side credential and persistent single-use guard; Vercel inference is not enabled. Tavily retrieval is implemented but has not been executed in the demonstrated workflow.

Observed runtime evidence

The completed recovery receipt records HTTP 200, response ID chatcmpl-b1ffe3497e224cc7, the matching Nemotron model, finish_reason stop, 263 prompt tokens, 262 completion tokens, and 525 total tokens. Guard elapsed time was 1,860.576 milliseconds, including the post-response checkpoint write; initial reservation and final terminal persistence are outside that timed span. This is not pure model latency. The token-based cost estimate was USD 0.0003147 using displayed regional rates. Final invoice cost and taxes were not verified. These observations describe one request, not a performance benchmark.

The receipt identifies application release sha256:c2b57f4b3b69dc3a7554a84932b5e23db23af652546379d7bd681e28927e101c and guard sha256:49aff9f118883edfe9f34067cd548f92bc7f1aa848653b0d10263065f9af5e2f.

Challenges and what we learned

An initial request mistakenly used a service-account ID as the credential and returned HTTP 401. We preserved that consumed attempt and its unknown cost. After explicit additional owner authorization and scoped review, a separately guarded recovery request succeeded. The recovery checks the retained prior failure record, preserves it, and prevents automatic retries. Masked dashboard text and service-account IDs are now rejected before reservation.

The key lesson was that a successful provider response, a matching quote, factual truth, and a verified business outcome are different claims. The interface and receipts must keep those distinctions clear.

Validation

Author-run offline checks passed 48 guard tests, 16 app/backend tests, and 22 JavaScript VM/loopback interaction assertions using fake transport and temporary records. A supplied external AI review reports reproducing those checks and returning a scoped GO for the exact recovery candidate. These are separate from the later owner-local live receipt. They do not constitute a native Safari, full accessibility, hosted tenant-isolation, or comprehensive security acceptance.

Feedback on Nebius and NVIDIA

We used Nebius Token Factory to run Nemotron’s assessment of supplied evidence. In the completed request, the API returned structured content, model identity, and token usage, and Nemotron identified the missing year rather than accepting the whole claim.

Our onboarding difficulty was distinguishing a service-account ID, masked dashboard representation, and complete API secret. More prominent credential labels and copy guidance would improve that experience. The first 401 was our credential-entry mistake; we do not attribute it to an outage. We did not use Nebius AI Cloud compute in this workflow.

Based on this limited test, we would continue evaluating Token Factory and Nemotron for evidence-review workflows because the response supported an inspectable review decision. Broader reliability, concurrency, latency distributions, and invoice-level costs require further evaluation.

Limitations and next steps

Original source authenticity and publication date, factual truth beyond the supplied text, cross-source contradictions, and business outcomes remain unchecked. fully_verified remains false. Source and receipt hashes establish internal consistency, not independent external provenance. The completed single-use inference attempt cannot be repeated under its consumed authorization.

Next steps are broader authorized source retrieval, repeatable deployment with durable spending controls, more substantive evidence cases, and independent end-to-end acceptance. These are planned work, not implemented capabilities.

Hackathon-period work

During the submission period, we developed and refined the evidence-review interface, deterministic checks, receipt display, owner-local Nebius/Nemotron integration, single-use authorization and persistence controls, adversarial regression checks, and candidate-bound review packets. The demonstrated runtime evidence belongs to the exact recovery candidate identified above.

Built With

• javascript
• nebius-token-factory
• nvidia-nemotron
• python
• vercel