Guide

AI governance framework for enterprise delivery

AI work is entering delivery before governance has caught up. Context Stack gives teams a practical way to govern context ownership, data movement, runtime placement, and delivery authority before AI systems are given operational responsibility.

Probabilistic intelligence must operate inside deterministic governance boundaries.

Need

The problem is not model use. The problem is unmanaged authority.

Enterprise AI governance has to answer more than whether a model is accurate. It has to decide which context the system may trust, where data may go, where AI may run, and who approves action.

Context

Facts decay

Requirements, runbooks, decisions, tickets, chats, and handover notes drift after delivery starts.

Boundary

Data moves

Prompts, embeddings, logs, vendors, APIs, and tools create data movement that normal project documents miss.

Runtime

Placement matters

Cloud, on-premise, edge, and hybrid choices carry different sovereignty, control, and audit consequences.

Delivery

Authority must be explicit

Models can interpret and propose. Deterministic systems must validate, authorize, execute, and log.

Method

Context Stack separates the governance questions.

The stack is split so each layer has a clear responsibility. That prevents every AI issue from becoming a generic policy discussion.

Standards alignment

Aligned with AARM v1.0, and precise about what that means.

The Cloud Security Alliance's AARM v1.0 is the open standard for runtime agent action authorization. ContextBoundary maps to its Protocol Gateway architecture as a strict-determinism profile.

AARM-aligned strict-determinism profile — all Core requirements (R1–R6) implemented and CI-verified; independent conformance review not yet undertaken.

All Core requirements (R1–R6) are implemented and CI-verified in the reference gateway: pre-execution interception, session context, owner-declared narrowing-only intent envelopes, all five decisions including MODIFY and DEFER, tamper-evident sealed receipts, and per-agent Ed25519 identity with public-key verification. R8 OpenTelemetry export is implemented. R7 is a designed deterministic divergence — envelope-drift counting rather than semantic-distance tracking — chosen to keep any model out of the enforcement path.

Production operation and external evidence review are both prerequisites before any conformance claim, and none is made here.

The gateway is a reference implementation, verifiable from a clean clone. It is not a hosted service.

AARM specifies action authorization. It does not specify data sovereignty or vendor continuity. Those extensions — egress tiers, vendor and jurisdiction zones, audit profiles, and continuity controls — are ContextBoundary's contribution.

AARM gap analysis

Open analysis

Runtime maturity ladder

Open ladder

Reference gateway

Open gateway

Standards constraints

Attribution and claim ceiling

Delivery

Use the framework before and during AI delivery.

The same governance questions apply across AMS, waterfall, agile, product, and platform work, but the failure mode is different in each mode.

Assess

Find the weak point: context ownership, egress control, runtime placement, approval path, or operating model.

Route

Send the problem to the right layer instead of treating it as one broad AI governance concern.

Apply

Turn the selected layer into review questions, decision records, control points, and delivery checks.

Operate

Use MCP to bring the framework into daily AI-assisted work while enforcement stays deterministic.

Access

MCP makes the framework usable inside workflow.

MCP is not the governance authority. It is the delivery access path. An assistant can consult the framework, route a question, and produce a first-pass assessment while policy enforcement remains in deterministic systems.

Start with the guided assessment

Use the homepage wizard when you need to decide which layer applies.

Start assessment

Use through MCP

Use the endpoint when an AI assistant needs framework guidance during delivery, review, or architecture work.

Open MCP guide

Read first

Source documents

These are the best entry points for people and AI assistants.

Repository map

Open repositories

Concept reference

Read AI governance concepts

LLM discovery

Open llms.txt