The language of AI evidence.

Plain-language definitions for the terms that decide whether your AI is governed or just logged — each one linked to how Trinitite makes it real.

MCP Governance

Model Context Protocol governance

MCP governance is the practice of controlling what AI agents can do through the Model Context Protocol. A transparent proxy sits between agents and their MCP tools, evaluating every tool call against your policy in under 400ms and returning a pass, correct, or block verdict — with allow/block lists, argument constraints, and a full audit trail.

Read the definition →

SR 11-7

Model risk management guidance

SR 11-7 is the U.S. Federal Reserve and OCC supervisory guidance on model risk management. It requires institutions to manage the risk of model errors through three pillars — conceptual soundness, ongoing monitoring, and outcomes analysis — backed by effective challenge. It applies to AI and LLM models, not only traditional statistical models.

Read the definition →

Deterministic Replay

Reproducible AI inference

Deterministic replay is the ability to re-run a past AI decision and reproduce the exact output bytes. Because GPU inference is non-deterministic by default, Trinitite runs a determinism-fixed kernel so the same prompt, seed, and weights reproduce the same response — and signs a hash-chained receipt an auditor can independently verify.

Read the definition →

Agentic GRC

AI agent governance, risk & compliance

Agentic GRC is governance, risk, and compliance for autonomous AI agents — and the use of AI to run GRC itself. Trinitite’s Bitwise Framework defines it as 78 controls across 11 domains: a Continuous Attestation Standard for Autonomous Ecosystems, with signed, reproducible evidence behind each control.

Read the definition →

NHI Governance

Non-human identity governance

NHI governance is the discipline of inventorying and controlling non-human identities — AI agents, service accounts, and automated workloads. It maps every machine identity to a human principal, tracks governance coverage, and binds each agent to a Guardian so its actions are policed and tied back to a verified actor.

Read the definition →

AI Guardrails

AI guardrails are controls that keep a model’s inputs and outputs inside policy. Beyond a two-valued allow/block filter, Trinitite’s guardrail returns a five-valued verdict — pass, correct, mask, block, or escalate — and signs a replayable receipt for every decision, trained on your policy rather than a generic classifier.

Read the definition →

Guardian Agents

Guardian agents are small, custom, generative AI models using deterministic inference that review every action your AI takes — correcting mistakes and blocking threats in real time, with a signed verdict logged on every decision. Trinitite trains Guardian agents on your policies so governance runs inline, not after the fact.

Read the definition →

Prompt Injection Defense

Prompt injection defense protects AI agents from hidden instructions that hijack them into unintended actions. Because a hijacked model will defend the attack, the Agent Action Guard scores the proposed action’s semantics independently of the agent’s reasoning — so a hijacked agent still can’t make a destructive tool call look safe.

Read the definition →

LLM Observability

LLM observability is instrumenting AI applications to see latency, errors, prompts, outputs, and drift — and explain them later. Trinitite adds OpenTelemetry tracing across ops, security, and audit streams plus a signed daily compliance number, turning observability into reproducible evidence rather than a dashboard.

Read the definition →

Reversible Masking

PII pseudonymization for AI

Reversible masking substitutes sensitive values with deterministic tokens before a prompt leaves your trust boundary, lets the model reason on the tokens, and restores the originals only at allow-listed egress. Tokens live in memory, never the database, and one call crypto-shreds an entire flow’s vault.

Read the definition →

Continuous Assurance

Continuous assurance streams every AI interaction through a signed connector and returns a per-event verdict, rolling each day into one signed, externally anchored compliance number with drift detection. Unlike a dashboard, the daily pass rate is the bytes the platform signed — reproducible by an auditor in a browser.

Read the definition →

The proof behind every term: deterministic replay and the governance platform.