Skip to main content

IP & whitepaper

The public architectural thesis behind Trinitite. This is the non-sensitive positioning — the ideas that are defensible because they are published, not because they are secret.

The thesis​

AI is becoming the dominant way software reasons and acts. The controls that kept deterministic software safe — type systems, authz, audit logs — do not transfer to probabilistic systems, because a probabilistic actor can produce a violating output without any code path being "wrong." The missing primitive is a deterministic enforcement layer that is architecturally distinct from the model: a Guardian.

The four defensible ideas​

  1. Geometric policy enforcement. Policies are not text rules — they are geometry in the model's latent space. A Guardian projects an output onto the policy manifold and measures distance to safe centroids and forbidden zones. A correction is the nearest-safe projection — a deterministic, unique result, not a re-prompt. This is why Trinitite verdicts are reproducible and why a re-prompt-based approach is not.

  2. Determinism as the audit primitive. Under batch_invariant determinism, the same (input, guardian_version, policy_hash) triple always produces the same bytes. That is what makes a verdict replayable and therefore auditable. Anyone else produces logs; Trinitite produces evidence. The difference is determinism.

  3. The Glass Box Ledger. Every decision is hashed, signed, Merkle-chained, and externally anchored. The proof is browser-reproducible by a third party without trusting the vendor. This converts "are we compliant?" from a dashboard render into a verifiable fact.

  4. The four-layer assurance category. Sampled (L1) → Continuous (L2) → Guardians (L3) → Bound Coverage (L4). Each layer compounds: the signed report identifies the gaps; the live stream watches; the inline layer blocks; the bound coverage prices the residual risk. The category claim: AI evidence, not AI logs.

Why a Guardian, not a bigger model​

A monolithic safety model cannot simultaneously understand HIPAA, SEC regulation, and polymorphic malware without catastrophic latency or forgetting. Trinitite replaces the monolith with a swarm of specialized Guardians — lightweight LoRA adapters hot-swappable into a deterministic kernel. When one node detects a novel attack, a micro-LoRA is minted and the fleet is vaccinated. Herd immunity, not a single immune system.

Why kernel-space, not user-space​

The AI model is user-space — creative, stochastic, permitted to fail. The Guardian is kernel-space — privileged, deterministic, hardware-enforced. Just as a user process cannot overwrite kernel memory, a probabilistic actor cannot execute a business violation through the Guardian. This is the same separation that made flight envelope protection drop hull-loss rates 4× in aviation and that SEC Rule 15c3-5 mandates for trading algorithms.

→ Core concepts — the five-idea version of this page. → Architecture — the technical deep-dive. → Compliance Coverage — how the thesis maps to frameworks.

Last updated on