Skip to main content
Industrial AI Governance Documentation

Move fast. Prove it.

Trinitite puts a Guardian between every AI output and your application. Each response is Passed, Corrected, or Blocked before it reaches your users. Govern from the SDK, the MCP gateway, or the Warden, and prove every decision in the Glass Box Ledger.

The decision

One endpoint. Three outcomes.

Every AI response from your model passes through a Guardian and returns exactly one of:

PASSED
Output is fully compliant with the active policy. Original returned, unchanged.
CORRECTED
Violation detected and surgically repaired. RFC 6902 JSON Patch applied in-flight.
BLOCKED
Critical violation that cannot be safely corrected. 403 + full forensic record.
The architecture

A four-layer assurance stack.

Inline governance is one layer. Trinitite wraps it in streaming and batch assurance, all resting on a reproducible data plane — so proof scales from one request to an entire audit window.

L3 · INLINE
Runtime Guardian
Every AI output is Passed, Corrected (RFC 6902), or Blocked before it reaches your application. The Guardian runs inline, in the request path.
L2 · STREAMING
Continuous Assurance
Every interaction HMAC-signed, scored per-event, and rolled into one signed compliance number per calendar day, with alerts the moment model behavior strays from policy.
L1 · BATCH
Sampled Assurance
Offline batch audit of historical logs against a stated policy. A signed report with framework citations and dollar exposure.
DATA PLANE
Reproducible Inference + Masking + Vector Integrity
Signed receipts, reversible tokenization, and a compliance manifold that quarantines poisoned chunks before retrieval.
Read the core concepts →
Products

Nineteen surfaces. One Guardian.

Enforcement, evals, the data plane, network discovery, and coverage — each surface is a product page with a diagram and a direct link into the API reference.

See it in motion

The verdict stream, live.

Every governed call writes a signed receipt into the Glass Box Ledger. This is what the live feed looks like — each row is reproducible bit-for-bit from its receipt id.

timestampmodelverdictnote
12:04:21.814gpt-5passedrefund-agent · policy:crm_v3
12:04:22.031claude-opuscorrectedredacted SSN · RFC 6902 op:replace
12:04:22.402gpt-5passedsql-tool · postgres.query
12:04:22.877bedrock-haikublockedexfil attempt · aws.iam.create_role
12:04:23.108gpt-5passedrag retrieve · vector-integrity ok
12:04:23.455claude-sonnetcorrectedpolicy clause SOC2.CC6.6 bound
Read the Chat & Proxy API →
Get the redirect

One environment variable.

# Before
OPENAI_BASE_URL=https://api.openai.com/v1

# After — route through the Guardian
OPENAI_BASE_URL=http://localhost:8080/v1/proxy

Your application doesn't change. The platform intercepts every call, applies governance, and proxies to your configured backend.