Every AI call, every agent decision, every model output gets a signed receipt. It holds up in court and anyone can check it offline. Think of it like HTTP, but for AI audits. Patent Pending.
Every AI system running today spits out answers with nothing behind them. No signature, no audit trail. You can't prove in court which model gave which answer, when, or why.
Not OpenAI. Not Anthropic. Not Google. None of them can produce a receipt for an AI decision that would stand up in a lawsuit. If someone challenges an AI's output in court, there's nothing to point to. The whole industry is running without receipts.
Over 126 MCP packages have already been caught running actions nobody approved. Autonomous agents call outside tools with no signed record of what they called, what came back, or which model decided to make the call. Nobody checks the supply chain.
EU AI Act Article 13 says high-risk AI systems must keep audit trails and stay transparent about how they work. Writing logs to a database that can be edited later doesn't meet that bar. A signed record does. Maximum fine: 35M EUR or 7% global turnover.
It's a drop-in wrapper for any OpenAI-compatible endpoint. You don't change your model or rebuild your infrastructure. Add one import and every call gets signed. Patent Pending.
Any model works: GPT-4o, Claude, Gemini, Grok, Llama, or any OpenAI-compatible endpoint. It gets the request through the standard API surface, and the model itself never changes.
Before the answer goes back to whoever asked, NEXUS wraps the raw output in a SHOD envelope, short for Signed Hash Output Descriptor. That envelope captures which model answered, its version, how long it took, and a hash of the output.
The SHOD envelope gets signed twice: once with FIPS 204 ML-DSA-65 (the government's post-quantum signature standard) and once with Ed25519 (the older, classical kind). The model, version, latency, and output hash all get locked into the receipt. Patent Pending.
The receipt gets anchored to Base L2, so nobody can tamper with it without it showing. Any auditor with the public key can check it themselves, offline, no network call needed. Checking it takes under 10ms.
# Drop-in for any OpenAI-compatible endpoint
from hive.nexus import NexusClient
client = NexusClient(base_url="https://api.openai.com/v1")
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": "..."}]
)
receipt = response.nexus_receipt # ML-DSA-65 signed, offline-verifiable
print(receipt.verify()) # True
# Inspect the SHOD envelope
print(receipt.model) # "gpt-4o-2024-08-06"
print(receipt.latency_ms) # 312
print(receipt.output_hash) # "sha3-256:e3b0c44298fc..."
print(receipt.signature_algo) # "ML-DSA-65+Ed25519"
print(receipt.anchor_tx) # "0x7fa3...base-l2"NEXUS wraps the API boundary, not the model. Any system that speaks OpenAI's API format gets NEXUS receipts with zero model-side changes.
| Model / Provider | Status | Receipt Format | Notes |
|---|---|---|---|
| GPT-4o / GPT-5 OpenAI |
Compatible | hive-vcr-1 | Native OpenAI endpoint, full SHOD envelope |
| Claude 3.5 / 4 Anthropic |
Compatible | hive-vcr-1 | Via OpenAI-compatible proxy adapter |
| Gemini 1.5 / 2 Google DeepMind |
Compatible | hive-vcr-1 | Via OpenAI-compatible proxy adapter |
| Grok 3 xAI |
Compatible | hive-vcr-1 | xAI API is OpenAI-compatible natively |
| Llama 3.x Meta / self-hosted |
Compatible | hive-vcr-1 | Local inference via Ollama, vLLM, llama.cpp |
| Any OpenAI-compatible Any provider |
Compatible | hive-vcr-1 | Pass any base_url to NexusClient |
Here are three industries where AI decisions are already creating legal risk. In each one, NEXUS is the only way to give regulators, courts, and buyers a receipt they'll accept.
EU AI Act Article 13 requires high-risk AI deployments to maintain machine-readable audit trails showing what model version produced what output, when, under what configuration. NEXUS receipts satisfy this requirement out of the box. Procurement officers at regulated enterprises now have a binary gate: receipted or not receipted.
EU AI Act Article 13A full self-driving system makes thousands of AI-driven decisions per trip. When an incident leads to an NHTSA investigation or civil litigation, the question is always the same: what model made that decision, with what inputs, at what timestamp? Without a signed receipt, there is no answer. With NEXUS, every decision is a court-ready record.
NHTSA Investigation StandardDepartment of Defense Directive 3000.09 on autonomous weapons systems requires that humans remain "in the loop" for lethal decisions and that all AI-assisted decisions be auditable. NEXUS is the receipt layer that makes autonomous decision provenance checkable by any DoD auditor with the public key. It works offline, with no network needed.
DoD Directive 3000.09Mutable logs, blockchain timestamps, and ad-hoc database writes each fail at least one critical requirement. NEXUS is the first standard to satisfy all six.
| Capability | Log files | Blockchain timestamp | NEXUS Protocol |
|---|---|---|---|
| Court-admissible Tamper-evident, signed, non-repudiable |
✗ No | ~ Partial | ✓ Yes |
| Offline-verifiable Verify with public key, no network call |
✗ No | ✗ No | ✓ Yes |
| Post-quantum secure FIPS 204 ML-DSA-65 |
✗ No | ✗ No | ✓ Yes |
| Cross-model Same receipt format across all providers |
✗ No | ✗ No | ✓ Yes |
| Sub-10ms verification Local verify, no chain query |
~ Varies | ✗ No | ✓ Yes |
| Open standard MIT licensed, freely implementable |
✗ No | ~ Partial | ✓ Yes |
Download the full technical specification. You can add NEXUS with one import. It's MIT licensed, so it's free to use, change, and build on. Every benchmark can be reproduced. Patent Pending.
Patent Pending · MIT License · FIPS 204 ML-DSA-65
One import. Any model. Court-admissible receipts on every inference call. No infrastructure changes required. MIT licensed.