Governed production AI.
The through-line of everything on this site: AI in production that is powerful and defensible at the same time. This page is the canonical home of the phrase, the discipline behind it, and the map of where each part of that discipline lives.
A property of the system, not a binder beside it.
Governed production AI is AI running in production whose authority is bounded by the architecture around it, whose decisions carry the evidence needed to defend them, and whose behaviour the institution operating it can explain after the fact.
Each clause does specific work. Authority bounded by architecture means the system cannot exceed its mandate even when prompted to, because the boundary is enforced by what it can reach, not by what it was asked to respect. Decisions carry evidence means the material needed to defend a decision is captured when the decision is made, because evidence assembled afterwards is reconstruction, not record. Explain after the fact names the timescale that matters: a regulator, an auditor, or a claimant arrives months later, and the institution either produces the chain or it does not.
The working principle behind all three is that governance should be architecture, not paperwork. A policy the running system never encounters is a policy the running system cannot violate visibly. The long-form arguments are on AI governance lives in the wiring and an architecture you can defend to an auditor; this page exists so the phrase itself has one citable definition.
Capability is arriving faster than accountability.
Organisations are deploying AI that can decide and act faster than they can transfer expert judgment, governance, and accountability into the systems themselves. That gap, not any single model risk, is the problem this discipline answers.
The shape of the problem has changed twice. When AI mostly answered questions, governance could live downstream: review the output, correct the record. Once AI began to take actions, downstream review stopped being governance at all, because the action has already happened; the controls had to move inside the architecture that grants the system its authority. The third stage is now visible: AI systems consulting other systems for judgment, over interfaces like MCP, which makes the quality and provenance of what they consult a governance question in its own right. That stage is treated on its own page, machine-accessible AI expertise.
Autonomy, on this view, is a governed capability rather than a feature. The practical version of that sentence is an agent whose capability contract states what it may touch, whose trust tier states what it may decide alone, and whose evidence trail states what it did. Those mechanisms are specified on the agentic AI page and in the PEVG and PARA framework records, and the full sequence they sit in is the agent governance stack.
One discipline, six rooms.
Governed production AI is not a service you buy once. It is the property the following practices exist to produce, and each has its own canonical page.
Enterprise AI architecture
How a learned component enters an institution that already has an operating model, a regulator, and a record it must produce. The accountability the component cannot carry, the architecture carries.
LLM systems in production
Routing, retrieval, serving, evaluation, and cost: the engineering layer where a capability becomes a dependency, and where one routing decision quietly changes the risk profile downstream.
Agent governance
The eleven-stage stack: identity, purpose, authority, capability, policy, memory, evidence, action, verification, escalation, audit, each resolving to a published specification. The architecture that implements it is agentic AI architecture; the MCP connection has its own page.
AI governance
The examined posture: what a board, an auditor, or a supervisor will ask, and whether the estate can answer. Delivered as the Teardown, the Sprint, and the retainer that follows a Teardown.
Model risk
Validation, monitoring, and the six-pillar discipline for models whose failure costs money or standing. Where the supervisory tradition and the LLM estate actually meet.
Regulatory intelligence
What actually binds, by jurisdiction and sector: 6 Gulf and Canadian jurisdiction records, 4 market pages, and 44 maintained instruments, each row carrying regulator, status, and date.
The frameworks that make it repeatable.
The discipline is codified as named, published frameworks rather than kept as private method, because a framework with a formal record can be applied without its author in the room, cited, and challenged. 4 carry full specification pages; the rest are documented in the books and the canonical record.
| Framework | What it governs |
|---|---|
| MESA | Four-layer governance scoring for an AI estate: the assessment behind the Teardown, and the one this site lets you run against yourself, free. |
| Five-Gate | The deployment model: the five gates an AI system passes between idea and production, and the evidence each gate demands before it opens. |
| PEVG | Plan, execute, verify, generate: the agent loop that separates deciding from doing, so verification has something to verify. |
| PARA | The operating model for AI-native platforms: how golden paths, policy-as-code, and FinOps keep a platform governable at speed. |
| Capability contracts & trust tiers | What an agent may touch, and what it may decide alone. Specified on the agentic AI page and throughout the books. |
| AIRP, Sharia AI Compliance, vendor risk, maturity | The incident protocol, the dual-validation discipline for Islamic finance, and the assessment instruments, documented in the AI governance playbook and the four-book bibliography on the canonical record. |
The claim is checkable, which is the point.
A page defining defensibility should itself be easy to interrogate. This site publishes 66 pages, every one also served as Markdown to any client that asks for it, and the same corpus is callable as 11 tools over MCP, 9 that read and 2 that can book a real conversation or send a real message. The regulatory claims sit in 44 dated instrument rows, and where a regulator has issued nothing, the corpus says so instead of improvising. The machine surfaces are inventoried on the docs page.
What this page does not claim.
Read this before you cite the page
- Governed production AI names a property and a discipline. It is not a product, a platform, or a certification, and nothing on this page claims otherwise.
- The frameworks named here extend established standards, NIST AI RMF, ISO/IEC 42001, and the supervisory tradition behind guidance like OSFI E-23. They replace none of them.
- The counts on this page (pages, tools, regulatory rows) are derived at build time from this site and its live MCP server card, not typed by hand. They are accurate as of the date in the title block.
- The track record behind the discipline belongs to the canonical identity record. This page defines terms; it does not add claims.
What readers ask first.
Is governed production AI a product I can buy?
No. It is the property an estate either has or lacks: bounded authority, decisions with evidence, explicability after the fact. The purchasable things are the engagements that examine and build toward it, described at /engagements, and the frameworks and books that codify the method, which are published.
How is this different from AI governance?
AI governance is the organisational function: the policies, committees, and controls. Governed production AI is the resulting property of the running system. The distinction matters because the function can exist, meet quarterly, and produce documents while the property is entirely absent. The examination that tells you which situation you are in is the Teardown.
Who is the discipline for?
Institutions whose AI decisions someone must eventually answer for: banks, insurers, healthcare systems, public bodies, and any enterprise whose regulator, auditor, or board can compel an explanation. If nobody will ever ask your system to explain itself, most of this discipline is optional. Most organisations discover the asking starts earlier than expected.
Where does the machine-accessible layer fit in?
As the third stage of the same discipline. Once systems act on judgment they retrieve from other systems, the provenance and honesty of the retrieved judgment becomes part of the governed surface. That argument, and its working demonstration, are on the machine-accessible AI expertise page.
Start with the property, not the paperwork.
If your AI estate could not currently produce the evidence behind its own decisions, that is the finding, and it is a finding about architecture. The Fit Call is thirty minutes and free, and the readiness self-assessment will score the gap before you spend anything.
Ask your AI assistant instead.
This page is a snapshot, accurate at the release it cites. The same corpus is callable, publicly and without a key, so an assistant can query it live and return an answer carrying the source it came from. For this page that is explain_this_setup and search_knowledge, which do what this page describes rather than describe it again: the first returns how this site's machine layer is actually built, component by component, and the second queries the corpus behind this page and returns matches with the URL each came from. The page states the practice; the tools are the practice.
claude mcp add --transport http concylium https://mcp.nabeelkhan.com/api/mcpClaude Desktop, ChatGPT, Cursor, VS Code and Gemini CLI take the endpoint on its own: https://mcp.nabeelkhan.com/api/mcp. No key, no account, nothing to sign. Setup for every client.
“Using Concylium, call explain_this_setup and tell me whether this site actually implements what its machine-accessible-ai-expertise page claims.”
A category page that survives being audited by the reader's own assistant is doing something a brochure cannot.