Verify independently
Re-execute a receipt on independent conforming hardware and compare hashes. A match corroborates the claim exactly; any differing byte proves a disagreement and triggers referee escalation.
Developers
Evidence with every response.
Mantissa exposes an OpenAI-compatible API. Change one base URL and every inference response arrives backed by a signed receipt: a falsifiable claim about exactly which model processed exactly which input, checkable by anyone, disputable on-chain.
The API
One base URL away.
If your stack speaks the OpenAI API, it already speaks Mantissa. Point your client at the gateway, keep your request format, and read the receipt alongside the completion.
Behind the gateway, jobs route to stake-bonded nodes that passed the conformance gate; results settle as receipts in epoch rollups. None of that changes your integration.
# Illustrative. Final shape ships with the public API docs.
curl https://api.mantissa.network/v1/chat/completions \
-H "Authorization: Bearer $MANTISSA_API_KEY" \
-d '{
"model": "qwen2.5-32b-verified",
"messages": [{"role": "user", "content": "..."}]
}'
# Response includes the completion, plus:
"receipt": {
"job_id": "0x91f3...",
"model_hash": "sha256:7c1e...",
"input_commitment": "sha256:b04a...",
"output_commitment": "sha256:5811...",
"checkpoint_hashes": ["04436b5a...", "..."],
"node_signature": "ed25519:..."
}The Receipt
What you can do with it.
A receipt is not a log line. It is a cryptographic commitment the network stands behind with stake. You, your customer, or an independent reviewer can check it.
One response. One receipt. The evidence travels with the output instead of living behind a provider dashboard.
Audit years later
Receipts are committed in epoch Merkle rollups on the settlement chain. A decision made today can be re-verified bit-for-bit long after the fact by you, a customer, a regulator, or a court.
Dispute mechanically
If a receipt is wrong, you do not argue about confidence scores. You challenge it. The protocol assigns conforming referees to re-execute the disputed work and compare exact hashes. Confirmed incorrect work enters the penalty path. Truth is a reproducible computation.
Is Mantissa For You
A premium product, honestly priced.
Mantissa is not the cheap option for commodity traffic. Verified consumer-GPU inference cannot beat centralized price floors, and we say so in the whitepaper. It is built for workloads where the output carries authority.
Strong fit
- Agent pipelines that move money, sign transactions, or modify production systems, with per-step receipts
- Compliance, adjudication support, and published research that may face challenge
- Reproducible ML: identical results on conforming GPUs, cross-vendor fault detection
- Session-heavy agent loops, where session affinity keeps context cached on your node
Honest non-fit
- Cost-driven bulk traffic where a receipt adds nothing; centralized commodity APIs will beat us on price
- Workloads needing cryptographic input privacy; serving nodes can read the prompts they process, and we will not pretend otherwise
- Frontier-scale closed models; the network serves open-weight models from its whitelisted registry
Closed Alpha
Tell us what you would verify.
We are onboarding a small set of design partners whose workloads exercise the receipt path for real. If that is you, we want to hear the specifics.