Security

TramAI Security provides a layered set of controls for production AI deployments: deny-by-default policy enforcement, data loss prevention, human-in-the-loop approval gates, build-time artifact verification, audit-grade evidence generation, and fully air-gapped runtime support.

Why This Exists

Production AI systems do more than call a model. They move sensitive context across provider boundaries, invoke tools, resume long-running workflows, and create decisions that may need to be justified to security, legal, compliance, or customer teams.

Sovereign mode exists to make those controls explicit instead of application-specific:

restrict where classified data can be routed
allow only approved models, providers, tools, and permissions
require human approval for high-risk tool execution
verify local model artifacts before the runtime starts
produce tamper-evident audit and evidence artifacts
enforce offline or air-gapped deployment rules at build time

The practical goal is simple: make governed AI behavior repeatable, testable, and reviewable before it reaches production.

The security surface spans two packages:

tramai-security — SPI definitions and implementations for DLP, audit, approval coordination, and artifact verification
tramai-sovereign — an aggregator module that composes tramai-security + tramai-standalone into a secure-by-default embedded runtime profile

For the detailed release-history view, see 0.4.0 — Sovereign Runtime and Governed AI Operations. It explains what changed between 0.3.1 and 0.4.0: the sovereign runtime, deny-by-default policy layer, trust-zone routing, approval gateway, hash-chained audit, evidence packs, artifact verification, offline deployment checks, and Spring Boot sovereign operations.

Feature Overview

Feature	Module	Complexity	Status	Use When
Sovereign Mode	`tramai-sovereign`	Medium	Stable	You need a secure-by-default runtime with deny-all policy, model allowlists, and trust zones
DLP	`tramai-security` + `tramai-core`	Low	Stable	You need to redact sensitive text (PII, secrets) from model outputs or tool results
Approval Workflows	`tramai-security` + `tramai-engine`	High	Experimental	You need human-in-the-loop approval before tool execution or workflow step resume
Artifact Verification	`tramai-security` + `tramai-core`	Medium	Stable	You need to verify model artifact integrity at build time (SHA-256, manifest enforcement)
Evidence Packs	`tramai-sovereign`	Low	Experimental	You need auditable, deterministic JSON evidence of deployment security posture
Offline Deployment	`tramai-sovereign`	Medium	Experimental	You need a fully air-gapped runtime with zero egress

Module Dependencies

tramai-sovereign
  ├── tramai-standalone
  ├── tramai-security
  │     ├── tramai-core (DLP SPI, ModelArtifactVerifier)
      └── tramai-engine (approval resume, SuspendedInvocationStore)

Quick Comparison

Concern	Sovereign Mode	DLP	Approval	Artifact Verification	Evidence Packs	Offline
Policy enforcement	Deny-by-default	—	—	—	—	—
Sensitive data redaction	—	Regex-based	—	—	—	—
Human-in-the-loop	Requires approval coordinator	—	Suspend/resume lifecycle	—	—	—
Build-time model integrity	—	—	—	SHA-256 streaming	—	—
Audit trail	Hash-chained audit engine	Redaction audit bridge	Lifecycle audit events	Verification receipts	Evidence packs	Zero-egress probes
Air-gap validation	Provider trust zones	—	—	Local-only verification	Zero-egress subsection	Full offline profile

Getting Started

The quickest path to a secure TramAI deployment:

// 1. Define your sovereign profile
val profile = SovereignProfileConfiguration(
    allowedModels = setOf("llama3.2"),
    allowedProviders = setOf("ollama"),
    providerZones = mapOf("ollama" to ProviderTrustZone.LOCAL),
)

// 2. Build the sovereign runtime
val tramai = SovereignTramai.builder()
    .profile(profile)
    .modelRegistry(registry)
    .auditStore(auditStore)
    .provider(ollamaProvider, name = "ollama", default = true)
    .model("llama3.2", "ollama")
    .build()

See the individual guides for each feature.

Next Steps

Sovereign Mode — secure-by-default embedded runtime
DLP — data loss prevention configuration
Approval Workflows — human-in-the-loop gates
Artifact Verification — model integrity checks
Evidence Packs — deployment attestation
Offline Deployment — air-gapped operations