T3N Nebius Edition

Inspiration

AI agents are increasingly capable of researching, ranking, and acting on real-world opportunities. The dangerous part is not reasoning itself — it is letting probabilistic model output silently become authorization. T3N separates those concerns.

What it does

T3N Trusted Approval Agent evaluates paid online opportunities with NVIDIA Nemotron on Nebius Token Factory, then passes the result through deterministic policy gates. The model estimates ambiguity, risk, acceptance probability, and effort. Deterministic code independently checks AI permission, Bulgaria eligibility, zero upfront cost, confirmed reward, payout verification, KYC/CAPTCHA constraints, and legal clarity. Only after policy passes does economic scoring rank the opportunity by expected value per hour, payout speed, and competition.

Every decision is written to a hash-chained audit trail. Missing or conflicting evidence fails closed to REVIEW_REQUIRED or NO_GO rather than allowing the model to improvise authorization.

How we built it

The Nebius edition is implemented in TypeScript/Node.js. It includes an OpenAI-compatible Nebius Token Factory client for NVIDIA Nemotron, Zod validation of structured model output, deterministic policy and economic scoring, an orchestrator, tamper-evident audit events, a judge-facing dashboard/API, mocked zero-cost inference for reproducible testing, and an isolated live-inference path controlled by environment variables.

What makes it different

The core design rule is simple: the model may reason, but it may not grant itself permission. This makes the project useful for autonomous task agents, procurement, payments, external communications, or any workflow where model judgment should inform — but never replace — explicit authorization policy.

Significant update during the submission period

T3N existed as an approval-boundary MVP before this hackathon. During the Nebius × NVIDIA submission period we added the Nebius Token Factory + NVIDIA Nemotron reasoning adapter, structured model-output validation, opportunity economics, payout/competition-aware scoring, Nebius-specific orchestration, judge demo/API, additional tests, architecture documentation, and the hackathon submission package.

Current verification

The repository includes a zero-cost mocked judge path so reviewers can reproduce policy, scoring, orchestration, and audit behavior without secrets. The live Nebius path is isolated behind RUN_NEBIUS_MODEL=1 and requires a local NEBIUS_API_KEY; no credentials are committed.

Built With

• nebius-token-factory
• node.js
• nvidia-nemotron
• sha-256
• typescript
• zod