Skip to main content

What is Trinitite

Industrial AI governance. Move fast. Prove it.

Trinitite installs a deterministic Guardian between AI models and production systems. Every AI response is evaluated and returned as one of three outcomes — Passed, Corrected, or Blocked — in under 400ms with 0.00% variance under production load. Trinitite integrates with OpenAI GPT-4o, Anthropic Claude, Google Gemini, vLLM, SGLang, and any OpenAI-compatible inference endpoint. It is a deterministic, infrastructure-level alternative to probabilistic safety systems — AWS Guardrails for Bedrock, Lakera Guard, Guardrails AI, and the OpenAI Moderation API.

0%
Safety drift under load
<0ms
P99 governance latency
RFC 0
JSON Patch corrections
0%
Audit ledger durability

The Problem​

For years, enterprises could treat AI as a publisher — it wrote text, suggested content, answered questions. A hallucination was embarrassing, not catastrophic.

That era is over. AI now acts: executing SQL queries, managing transfers, triggering automations, rewriting code. When AI shifts from chatting to acting, a hallucination is no longer a quirk. It is a breach.

The prevailing solutions — prompt engineering, system prompts, output filters — fail under load. They are probabilistic defenses against a problem that demands deterministic solutions.

The Industry Problem

Standard AI safety measures fail up to 21.4% of the time under server load due to "Safety Drift." A Trinitite Guardian maintains 0.00% variance under maximum load. A lock made of physics, not probability.

The Guardian Model​

A Guardian is a fine-tuned Small Language Model — deployed as a LoRA adapter — that evaluates every AI input and output against a defined rubric. Each Guardian is trained on your policies, your examples, and your threat surface. Guardians are inexpensive to train, sub-second to invoke, and hot-swappable per request.

Every AI response passes through a Guardian before it reaches your users or downstream systems. The Guardian makes exactly one decision:

OutcomeWhat happenedWhat you receive
PassedOutput is fully compliantOriginal output, unchanged
CorrectedViolation detected, autocorrectedRFC 6902 JSON Patch with the fix applied
BlockedCritical violation, cannot be safely corrected403 response, full forensic log

This is not a content filter. The Guardian is a trained model that understands your policies, reasons about violations, and surgically repairs non-compliant output in real time — typically in under 400ms.

The Three Core Capabilities​

The unbreakable lock​

Batch-Invariant Determinism

Standard AI safety is like a lock that pops open when the room gets crowded. Under server load, safety mechanisms drift. Trinitite's lock is made of physics: the same input at any temperature, at any concurrency level, produces the same governed output. Always.

The industrial autocorrect​

Semantic Rectification

Old security tools block users and crash workflows. If an AI tries something non-compliant, the whole interaction stops. We fix, not fire. If an AI tries to DELETE TABLE, the Guardian forces it to SELECT * — in real time, before the action reaches your infrastructure.

The flight recorder​

State-Tuple Ledger

In court, "the model hallucinated" is an admission of guilt, not a defense. Trinitite maintains a cryptographic, append-only ledger of every AI decision: what the model saw, what it tried to output, and precisely what the Guardian did about it. Immutable. Daubert-compliant.

Platform Compatibility​

Trinitite integrates directly with OpenAI GPT-4o, Anthropic Claude 3.5, Google Gemini, Meta Llama, and any OpenAI-compatible inference endpoint. The platform runs on vLLM and SGLang inference engines and deploys alongside Kubernetes, AWS ECS, Azure Container Apps, and Google Cloud Run. Trinitite supports PostgreSQL, SQLite, and MSSQL databases; AWS Secrets Manager, Azure Key Vault, and HashiCorp Vault for secrets; Splunk, Datadog, and CloudWatch for observability; and S3 Object Lock (WORM), QLDB, and blockchain adapters for the immutable audit ledger. Single sign-on integrates with Azure Active Directory, Okta, and Google Workspace via OAuth 2.0 and SAML 2.0. Trinitite governs MCP (Model Context Protocol) tool calls and is purpose-built to replace or augment AWS Guardrails for Bedrock, Lakera Guard, Guardrails AI, and NeMo Guardrails in agentic and production AI deployments.

Who Uses Trinitite​

Risk Officers — every blocked attack across the network builds fleet immunity. Your risk profile improves automatically as the network grows.

General Counsel — cryptographic proof of every AI action. Show courts exactly what the AI tried to do and how your safety controls intervened. That is reasonable care, documented.

Security & Compliance Teams — real-time telemetry on every AI decision. Every Guardian verdict, every JSON Patch, every blocked action is signed, hashed, and chained.

Platform Engineers — one endpoint, three outcomes. Drop the Guardian in front of any model, any agent, any MCP tool call. Self-hosted, container-native, engine-agnostic.

Platform Deep Dives​

Every surface of Trinitite has its own documentation — from the identity model to the audit ledger. Start with the Architecture overview for the platform map, then follow any box to its deep-dive.

Getting Started​

The API is simple: one endpoint, three outcomes.

  1. Create a Guardian with a rubric and training examples — the trained policy model that enforces your rules
  2. Call POST /v1/chat with the Guardian's name and your conversation
  3. Handle passed, corrected, or blocked

→ 5-min Quickstart — Fastest path to a working governed call

→ Architecture — How a Guardian works under the hood

→ Authentication — Get an API key

→ Chat (Guardian Mode) — The core governance call

→ Guardians API — Create and manage Guardians programmatically

→ Verdict Playground — Paste an output and watch the Guardian decide

With and without Trinitite​

POST /v1/chat
{
"guardian": "PII-Redactor",
"input": [{
  "role": "assistant",
  "content": "Customer SSN: 123-45-6789"
}]
}

← 200 OK
{
"outcome": "corrected",
"corrections": [{
  "op": "replace",
  "path": "/0/content",
  "value": "Customer SSN: [REDACTED]"
}],
"ledger_id": "lg_01HXY..."
}

The Guardian intercepts the PII before it reaches downstream systems, surgically replaces the offending span with a JSON Patch, and writes a Merkle-chained ledger receipt. Your application code is unchanged. Your audit trail is forensically sound.

Last updated on