SoverignLedger

Zero-trust contract verification, autonomous payout execution.

Comment

Inspiration

In global infrastructure, construction, and public procurement, over $120B is lost annually to administrative payment gridlock, contested change orders, and payouts for non-compliant equipment.

Today's milestone payouts still depend on manual human sign-offs. Contractors submit delivery manifests, field tickets, and inspection reports across disconnected PDF channels. Fatigued project managers often "rubber-stamp" approvals without verifying granular engineering standards, or they freeze payments for months over minor paperwork discrepancies.

We were inspired to build SovereignLedger: an autonomous AI referee operating on zero-trust principles to bridge dense legal contracts with unstructured physical proof and make milestone escrow payouts deterministic, immediate, and fair.

What it does

SovereignLedger is an autonomous forensic payment agent that continuously audits physical proof against legal contract clauses and controls escrow fund disbursement:

  • Parses & Indexes Contracts: Deconstructs complex legal contracts into machine-verifiable requirements, specifying exact technical codes (e.g., ASTM cement break ratings, UL fire certifications, and professional engineer licensing stamps).
  • Ingests Multi-Source Evidence: Ingests unstructured real-world proof—such as shipping manifests, bills of lading, and on-site inspection logs.

  • Forensic Arbitration: Performs zero-trust, multi-step legal and technical cross-examinations to ensure physical deliveries match contractual agreements down to the serial code.

  • Autonomous Escrow Control: If all conditions are met, it clears the milestone and generates an immutable cryptographic transaction hash ($0\text{x}\dots$) releasing the payout. If a specification mismatch or safety hazard is detected, it freezes the funds, holds the escrow, and generates an itemized clause violation report.

How we built it

We designed an end-to-end, heterogeneous compute architecture hosted on open infrastructure via Nebius Token Factory, pairing fast edge normalization with deep legal reasoning:

[ Unstructured Evidence & Contracts ]
                 │
                 ▼
     ┌────────────────────────┐
     │  nvidia/nemotron-nano  │ ◄── Nebius Token Factory (Extraction Tier)
     └───────────┬────────────┘     • 380ms latency, strict JSON normalization
                 │
                 ▼
     ┌────────────────────────┐
     │ nvidia/nemotron-3-ultra│ ◄── Nebius Token Factory (Reasoning Tier)
     └───────────┬────────────┘     • Deep legal cross-examination & arbitration
                 │
                 ▼
     ┌────────────────────────┐
     │  Escrow State Machine  │ ◄── Automated Payout / Arbitration Hold
     └────────────────────────┘
  1. Extraction Tier (nvidia/nemotron-nano on Nebius Token Factory): Acts as a high-throughput normalization layer, parsing raw document text, delivery tickets, and bills of lading into structured schemas in sub-400ms without schema leakage.
  2. Forensic Arbitration Tier (nvidia/Nemotron-3-Ultra-550b on Nebius Token Factory): Acts as the contract auditor, cross-examining every technical specification against legal covenants and outputting verifiable audit traces.
  3. Backend Orchestration (FastAPI & Python): Runs the arbitration pipeline, coordinates with Nebius Token Factory's OpenAI-compatible API, and computes verification hashes:

$$\text{TxHash} = \text{SHA-256}(\text{MilestoneID} \mathbin{\Vert} \text{Timestamp} \mathbin{\Vert} \text{AuditorVerdict})$$

  1. Cyber-Finance Dashboard (Next.js 14, TypeScript, & Tailwind CSS): Delivers a high-contrast terminal interface featuring milestone matrices, evidence injectors, and real-time streaming of Nemotron 3 Ultra's step-by-step reasoning trace.

Challenges we ran into

  • JSON Schema Enforcement Under Complex Reasoning: Frontier reasoning models often wrap structured output in explanatory markdown when performing complex multi-step deductions. We resolved this by pairing Token Factory's strict JSON response formatting with deterministic regex sanitization layers in Python to guarantee zero parser crashes.
  • Minimizing Latency for Dense Documents: Routing multi-page manifests directly into a massive reasoning model caused high inference latency. We re-engineered the pipeline into a two-tier model topology: delegating data structuring to nemotron-nano and reserving nemotron-3-ultra solely for high-stakes clause arbitration, cutting turnaround time by over $70\%$.
  • Edge-Case Nuance in Technical Standards: Distinguishing between closely related industrial safety standards (such as UL-1973 for stationary indoor storage vs. UL-9540 for utility-scale fire protection) required rigorous few-shot calibration in Ultra's system prompt to ensure zero false positives during audit verification.

Accomplishments that we're proud of

  • Heterogeneous Model Synergy: Successfully orchestrating Nemotron Nano and Nemotron 3 Ultra on Nebius Token Factory to achieve a $10.1\times$ speedup on extraction and a $95\%$ reduction in token costs compared to monolithic alternatives.
  • Deterministic Discrepancy Detection: Demonstrating a zero-failure live demo where our agent successfully intercepts a critical fire-safety substitution (UL-1973 vs. UL-9540), freezing a $\$75,000$ escrow transfer in real time.
  • End-to-End Live Product Experience: Building a functional platform that merges legal analysis, streaming reasoning traces, and automated settlement records into an intuitive user interface.

What we learned

  • Model Tiering is Critical for Enterprise AI: You do not need to route every byte of unstructured data through a massive reasoning model. Fast, lightweight models like Nemotron Nano are exceptionally competent at schema normalization, allowing larger models to concentrate purely on deduction and risk analysis.
  • Open Infrastructure Flexibility: Developing against Nebius Token Factory proved how accessible and responsive open-source model serving can be when built on standard, OpenAI-compatible interfaces.

What's next for SovereignLedger

  • Multi-Modal Vision Inspection: Integrating NVIDIA Cosmos and vision models to evaluate drone imagery, geotagged photos, and thermal camera feeds directly against blueprint specifications.
  • Smart Contract Integration: Deploying on-chain escrow smart contracts (e.g., Ethereum/Polygon or Solana) where Nemotron 3 Ultra signs off on programmatic liquidity pool releases via account abstraction.
  • ERP & Supply Chain Connectors: Building native webhook adapters for enterprise procurement systems (SAP, Procore, Oracle Aconex) to trigger forensic audits the moment shipping carriers scan a parcel at the job site.

Built With

Share this project:

Updates