The logic layer for regulated AI execution
An execution-control layer for enterprise AI, built on a clean architectural separation between the probabilistic and the deterministic.
What Luminous is
The language model — probabilistic by nature — interprets and proposes. A deterministic, sealed policy engine decides and executes. Every guarantee the system makes lives entirely on the deterministic side; the probabilistic model is contained and non-dispositive. No consequential action executes unless the deterministic conditions are satisfied. The organizing property is deliberately non-obvious: the AI holds no authority to act, by construction — not as a policy it can be talked out of. Regulated industries are the sharpest wedge, but the same engine governs any enterprise that needs controlled AI execution.
Governance and assurance
Governance — keeping AI on the rails — is the floor most enterprises need. Assurance — containing a model you must assume is hostile, and proving it — is the depth regulated and high-stakes environments require. Luminous delivers both from one engine, with assurance as a selectable dial.
Four properties
Deterministic and probabilistic components, cleanly separated
Nothing that must be exact, auditable, or reproducible depends on a model's probability.
Signed, versioned, auditable operating logic
Reconstruct exactly which rules, permissions, and definitions were in force at the moment of any decision.
Deterministic data governance
Every field carries an enforced classification; access is gated by signed policy, never model discretion.
Human authorization and a record that remains
A person authorizes each consequential action; every governed request emits a signed, tamper-evident receipt to a write-once audit fabric.
Deployment — one engine, three assurance tiers
Assurance is a property of the deployment; moving up a tier is a migration, not a re-architecture.
| Tier | Description |
|---|---|
| Assertion | Cryptographic integrity on standard compute. |
| Attestation | Proves it in hardware, on a confidential-compute substrate. |
| Sovereign | Removes the cloud provider from the trust base entirely. |
Proof, honestly stated
No reference logos, so we don't ask you to take our word. The enforcement gate has been reduced to practice on real hardware and put through an internal, separation-of-powers adversarial validation — an independent red-cell seat confirms the controls actually bite, with reproducible evidence a counterparty can re-run under NDA.
This is containment, not immunity; today's strongest validation is at the Assertion tier on synthetic data; model-diverse and human validation passes are named next steps, not claimed results. IP protected by a filed provisional and a comprehensive invention-disclosure record.
Prefer to read first?
Ask for the two-page brief, or set up a 20-minute call.