Without privacy, AI is surveillance.Without proof, AI cannot be trusted.Without determinism, AI cannot be audited.
Paradatum runs open models deterministically, privately and with proof, at full precision. The machine doing the work never reads your data. Every answer comes back with a checkable receipt.
AI inference you can hand to an auditor.
- Deterministic
- Private
- Proven
- Lossless
Same prompt, same model, three GPU nodes
SHA-256 of the output bits
Bit-exact match. Verification is a hash comparison.
Illustrative receipt.
Four guarantees, one engine.
For a restaurant recommendation, the gaps in ordinary AI are tolerable. For a loan, a claim, a contract, a diagnosis or a filing, they are not.
Deterministic: reproducible
One agreed computation, one canonical result, on any GPU (graphics processing unit). Fixed execution order, no floating-point drift.
Private: not surveillance
The operator running the model cannot read the prompt, the documents or the answer. Plaintext exists only inside your own trust domain.
Proven: verifiable
Every job returns a receipt and a proof that the agreed model ran on the committed inputs. A third party can check it later without re-running anything.
Lossless: full precision
Exact BF16 (16-bit brain floating point) weights with no quantization loss. Lossless compression gives 25% less VRAM (video memory) and up to 2× throughput.
For enterprises: accountable compute.
A regulated institution cannot defend a consequential decision with “the API returned this.” Underwriting, lending, compliance, legal review, investment analysis and risk decisions need a record. The institution needs to know:
- which model was agreed,
- which version ran,
- what computation was performed,
- what inputs were committed,
- whether the result can be reproduced,
- and whether a third party can verify the record later.
Paradatum returns that record with every job. The proof is portable: it outlives the job, the server and the vendor relationship. Because the operator never sees plaintext, the same job can run on a rented GPU, a partner’s datacenter or your own, with the same guarantee.
For the people inside them: private thought.
The most valuable question is often the one a person is least willing to type. What could this symptom mean. Can I afford this. Is my employer in breach of this agreement. Today, asking means handing the question to the company operating the machine.
The line between AI as a service and AI as surveillance is whether the service can do the work without reading the person using it.
With Paradatum the client masks locally, the compute happens elsewhere, the answer returns masked, and plaintext never leaves the user’s own trust domain.
How a job runs.
Five steps, one receipt chain. Nothing in the middle can read what it is computing.
You send a masked job
Your client masks the prompt and documents before anything leaves your trust domain.
Paradatum pins the contract
Model, version, precision and execution path are fixed and committed before the run.
A GPU runs it blind
Any supported hardware, yours or rented. The operator sees masked bytes only.
Answer, receipt, proof
The output returns with a hash of its bits, chained to the contract and the committed inputs.
Anyone can re-verify
Re-check the proof in milliseconds, without re-running the model, for as long as the record matters.
Performance
Measured against standard BF16 inference.
| Measure | Result | What it means |
|---|---|---|
| Throughput | Up to 2× | Bit-exact speed gains across platforms. |
| VRAM | 25% less | Lossless compression of full-precision weights. No quantization. |
| Reproducibility | 100% match | Identical output bits on every run and every GPU. |
| Audit | Hash-verified | Receipt and proof with every job; re-verify without re-running. |
| Privacy | Operator-blind | Plaintext never exists on the serving machine. |
| Portability | Any supported GPU | Consistent results across hardware. No vendor lock-in. |
See it run
A short walkthrough of deterministic execution and hash-verified outputs.
Open on YouTubeThe Paradatum family.
Paradatum is the gateway and the engine. The same guarantees power three open networks. Enterprise customers deal with Paradatum, pay in dollars, and never touch a network token.
Frequently asked questions
What does “private” mean here?
The machine running the model operates only on masked data. It cannot read the prompt, the documents or the answer. Plaintext exists only inside your own trust domain, on your client.
What does “proven” mean?
Every job returns a receipt: the agreed contract, the committed inputs, and a proof that the computation ran as agreed. A third party can check the proof without trusting Paradatum or the operator, and without re-running the model.
Why does determinism matter if I already have privacy and proof?
A proof binds to one canonical result. Without a deterministic execution path there is no single result to prove, and no way to reproduce a decision later. Determinism is what makes the other two guarantees possible.
What does “lossless” mean?
The model runs at full BF16 precision. The 25% memory saving comes from lossless compression of the weights, not from quantization, so answers are bit-identical to the uncompressed model.
Can it run on our own hardware?
Yes. The guarantees hold on your GPUs, a partner’s, or rented capacity. The operator of the hardware never sees plaintext either way.
How does this compare to INT8 or BF16 inference?
INT8 (8-bit integer) quantization trades accuracy for speed and gives up bit-exact reproducibility. Standard BF16 keeps accuracy but drifts from run to run and machine to machine. Paradatum keeps full BF16 precision, removes the drift, and adds privacy and proof on top.
Talk to us
For technical documentation, enterprise integration, a pilot, or support. We respond within one business day.