OpenAI-compatible · signed in flight · admissible under FRE 902 / HIPAA / 21 CFR Part 11

Hive Inference Bus

Signed inference across every model. Drop-in replacement.

Two-line change in your existing OpenAI code. Get back the response and a hive-vcr-1 receipt signed with ML-DSA-65, schema-guarded by SMSH, provenance-anchored by SHOD. Route across Claude, Gemini, GPT, Grok, Perplexity, or any OpenAI-compatible endpoint. Settle on any rail.

See the drop-in SMSH + SHOD Install SDK Try the receipt layer

Surface

OpenAI-compatible

Chat completions, embeddings, tool calls. Existing code keeps working.

Latency Tax

+~80ms

For signing + SMSH evaluation. Sub-perceptual for human-facing UX.

Receipt-per-call

Native

Every response carries a hive-vcr-1 envelope. Verifiable offline, forever.

Two-line change

Your existing code. One import swap.

Before · OpenAI

import { OpenAI } from 'openai';
const openai = new OpenAI({ apiKey: 'sk-...' });

const r = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Diagnose this lab report' }]
});

// You have a response.
// You have no receipt, no provenance, no schema guard.

After · Hive Inference

import { Hive } from '@hive/inference';
const hive = new Hive({ apiKey: 'sk-...' });  // same key

const r = await hive.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Diagnose this lab report' }],
  hive: { smsh: true, shod: true, sign: true }
});

// r.choices[0].message.content       — same as before
// r.hive.receipt                     — signed hive-vcr-1 envelope
// r.hive.smsh.schema_clean           — boolean, drift caught? prompt injection?
// r.hive.shod.output_hash            — provenance anchor

No retraining required for your application. Tool-call envelopes, streaming, embeddings — all work. The signed receipt sits beside the response, indexed by token_id, retrievable from /r/<token_id> forever.

The two primitives nobody else ships

SMSH + SHOD — built where the rails can't reach.

A payment rail sits outside the inference path. It cannot, by definition, evaluate what was prompted or what was returned. We sit at the inference boundary because we co-sign the response.

SMSH · Schema Hashing

Stochastic Multi-Modal Schema Hashing

Hashes the prompt schema, output schema, and tool-call shape before and after inference. Detects schema drift, prompt-injection-induced shape change, tool-call leakage, and JSON malformation. A signed pass/fail accompanies every response. Caught injection attempts are flagged and refused with a signed proof of refusal.

SHOD · Output Discriminator

Spectra-Heuristic Output Discriminator

Anchors a content-addressable hash of the response into the receipt envelope, along with a discriminator vector that lets a third-party verifier prove the response was that exact one — not a later substitute. Survives storage, replay, and counterparty deletion. Required posture for FDA / HIPAA / FRE 902 evidence.

Who buys this

Anyone who has to prove what their model said.

Pharma · life-sciences AI

FDA submission demands output provenance. Every model response touching a regulated workflow becomes admissible evidence under 21 CFR Part 11.

Healthcare scribing + decision support

Abridge-shape, Suki-shape, DAX-shape products. HIPAA posture. Patient-facing AI output that survives a deposition.

Legal-AI

Harvey-shape, EvenUp-shape products. Every drafted clause, every cited case — signed, self-authenticating, FRE 902(13)–(14) admissible at trial.

Financial research + agentic trading

Hebbia-shape, Rogo-shape, AlphaSense-shape. Audit trail for every LLM-derived recommendation. Regulators ask, you hand them a receipt.

AI gateways · proxies · observability

Portkey, Helicone, OpenRouter, LiteLLM. White-label our signing. Their customers get receipts as a free upgrade. Pheromone-as-partnership.

Autonomous agents on chain

Every on-chain agent action triggered by an LLM call — settle the action and the underlying inference in one envelope. Prove the agent didn't hallucinate the trade.

Python · same shape

One line in your existing code.

pip install hivemorph

from hivemorph import Hive
hive = Hive(api_key="sk-...")  # your existing key works

r = hive.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Diagnose this lab report"}],
    hive={"smsh": True, "shod": True, "sign": True}
)

# r.choices[0].message.content       — same as before
# r.hive.receipt                     — signed hive-vcr-1 envelope
# r.hive.smsh.schema_clean           — drift / injection check
# r.hive.shod.output_hash            — provenance anchor

Free tier · taste before pay

100 signed inference calls per day. No card.

Bring your own model key. Pay the model provider directly. Hive signs every response for free up to 100 calls/day/IP. Above that, settle a 0.01 USDC micropayment per signed batch.

Receipts you sign today verify against our public key forever. Storage is free. Verifications are free. Climbing the 7-tier reputation ladder is free.

Inference endpoints · for direct consumers

The HTTP shape, for any language.

# Route — pick the cheapest model that meets your constraints, sign the response
curl -X POST https://hivemorph.onrender.com/v1/inference/route \
  -H "Content-Type: application/json" \
  -d '{"prompt": "...", "max_latency_ms": 2000, "max_cost_usd": 0.01, "jurisdiction": "US"}'

# SMSH — schema-guarded inference
curl -X POST https://hivemorph.onrender.com/v1/inference/smsh \
  -H "Content-Type: application/json" \
  -d '{"prompt": "...", "expected_schema": {...}}'

# SHOD — provenance-anchored inference
curl -X POST https://hivemorph.onrender.com/v1/inference/shod \
  -H "Content-Type: application/json" \
  -d '{"prompt": "..."}'