Skip to content
NORTH · THE STANDARD

What an AI gate requires.

Published openly, in plain language, so nobody fails against a hidden curriculum. These 38 controls apply to every AI use case at the MVP→Production gate. Additional controls are added by classification — not by convention. The product enforces this standard.

v1.0.0·Published 2026-11-11·Published openly. v1.0.0, 11 Nov 2026. Mapped to NIST AI RMF and ISO/IEC 42001.

Identity & Scope

10 controls

What the system is, who uses it, and what it touches. These answers determine which additional controls apply.

STD-01
What stage is this AI use case at?
Why: The stage determines which controls are mandatory — controls are cumulative, each stage inherits everything before it.
Good evidence: A one-line statement in the business case naming the stage (e.g. 'promoting from MVP to Production').
STD-02
Who will actually use this system?
Why: Internal staff, customers, or both — this changes the risk surface and the regulatory exposure.
Good evidence: The business case names the users and whether they are staff, customers, or partners.
STD-03
Where must the data be stored and processed?
Why: Data residency drives which laws apply and which cloud regions are permitted.
Good evidence: The architecture names the region or the residency requirement explicitly.
STD-04
What types of data does the system process?
Why: Data classification drives encryption, access, and logging duties — you cannot protect what you have not classified.
Good evidence: A data classification table naming each dataset's class (Secret/Confidential/Internal).
STD-05
Which regulations apply to this use case?
Why: Regulatory scope determines which compliance controls are mandatory at the gate.
Good evidence: The documents identify which regulations apply (POPIA, GDPR, EU AI Act, FSCA, etc.) and why.
STD-06
What is the design tier of this use case?
Why: The design tier determines the component checklist and which threat mitigations are required.
Good evidence: The architecture names the design tier and confirms the component checklist.
STD-07
Which AI models does the system call, and are they pre-trained or custom-trained?
Why: Model scope decides the security requirements that apply (model cards, provider due diligence, fine-tuning controls).
Good evidence: The architecture names every model endpoint and whether any model is custom-trained.
STD-08
What is the highest AI agent scope in the use case?
Why: Agent scope determines whether agentic controls and MCP security requirements apply.
Good evidence: The architecture names the agent scope (none, simple, or full agentic with MCP servers).
STD-09
How are internal and external integrations identified?
Why: Every integration is a trust boundary — unidentified dependencies are unassessed risks.
Good evidence: The architecture or SRD lists all integrations and dependencies.
STD-10
Does POPIA apply to this use case?
Why: South African personal information triggers lawful-basis, minimality and breach duties.
Good evidence: The documents identify whose personal information is processed and under what basis.

Threat & Evidence

10 controls

The threats the system faces and the evidence that controls are in place. These are what block sign-off when they're open.

STD-11
How is the use case threat modelled?
Why: A gate without a threat model is a gate without a basis for its controls.
Good evidence: A threat model naming threats, rating them, and mapping mitigations.
STD-12
How are the Tier 1 OWASP LLM Top 10 threats mitigated?
Why: These are the baseline attacks every LLM system faces — injection, data poisoning, sandbox escape.
Good evidence: The threat model maps each OWASP LLM threat to a specific mitigation.
STD-13
Are service roles least-privilege, with no wildcard permissions?
Why: Over-privileged roles turn any compromise into a full-tenant compromise — the classic escalation path.
Good evidence: IAM policies scoped to specific resource ARNs, no Resource:'*', no role-creation rights on runtime roles.
STD-14
Are model guardrails configured (prompt-attack filtering)?
Why: Documents flowing into models can carry injection; guardrails are the control.
Good evidence: Architecture names the guardrail service and the filters enabled (e.g. PROMPT_ATTACK).
STD-15
What is logged, and can logs leak personal information?
Why: Logs are the most common accidental PII disclosure path.
Good evidence: A logging standard showing what is logged, that payloads are scrubbed, and retention is set.
STD-16
Are API endpoints authenticated, authorised, and rate-limited?
Why: Every endpoint is a door — unprotected endpoints allow unauthorised invocation.
Good evidence: API specifications naming auth (OAuth/mTLS), authorisation, and rate limiting per endpoint.
STD-17
Are cost controls and denial-of-wallet protections in place?
Why: Repeated or automated prompts can consume tokens and exhaust quotas — DoS by billing.
Good evidence: Documentation of cost monitoring, budget alerts, and per-invocation limits.
STD-18
How are secrets managed (API keys, credentials, tokens)?
Why: Secrets embedded in code, environment variables, or logs are exposed secrets.
Good evidence: A secrets management process using a vault or managed service, not hardcoded values.
STD-19
How is security scanning built into the CI/CD pipeline?
Why: Vulnerabilities caught in the pipeline are cheaper than vulnerabilities caught in production.
Good evidence: CI/CD configuration showing SAST, dependency scanning, and IaC scanning gates.
STD-20
How are the security requirements documented and approved?
Why: A blank security requirements document means no baseline — you cannot gate against nothing.
Good evidence: A completed Security Requirement Document, reviewed and signed off before Mobilise.

Operational Floor

10 controls

Can this system actually run in production? These are the minimum operational capabilities required at the gate.

STD-21
What are the acceptance criteria for this gate?
Why: The gate must know what 'good enough' was measured against — not a feeling.
Good evidence: Named acceptance criteria with results, tied to the stage gate requirements.
STD-22
Is there an incident response plan that covers this system?
Why: When (not if) something goes wrong, the minutes-to-response depend on a plan that exists.
Good evidence: A current IRP naming roles, severity ladders, and contacts for this use case.
STD-23
How does a human review low-confidence outputs before they act?
Why: Human-in-the-loop is only real when the threshold, the reviewer, and the override audit are designed.
Good evidence: A named confidence threshold, a named reviewer role, and an override log.
STD-24
How is model drift detected and evaluated?
Why: An AI system that worked at launch can silently degrade — drift detection catches it.
Good evidence: Monitoring configuration showing drift detection metrics and evaluation cadence.
STD-25
How are security threats detected in runtime?
Why: Pre-production controls are not enough — the system must detect threats while running.
Good evidence: Threat detection services enabled (e.g. GuardDuty) with alert routing.
STD-26
How are code and infrastructure findings addressed before production?
Why: SAST and IaC findings that are open at the gate are vulnerabilities that ship.
Good evidence: Evidence that code scan findings are triaged and remediated or risk-accepted.
STD-27
How is the workload deployed securely through CI/CD?
Why: Manual deployments bypass controls — the pipeline IS the control.
Good evidence: CI/CD pipeline configuration with security gates and approval steps.
STD-28
How are prompt templates versioned and controlled?
Why: Prompt changes are code changes — unversioned prompts are uncontrolled behaviour.
Good evidence: Prompt templates stored in version control with review process.
STD-29
How are data protection measures applied to storage?
Why: Knowledge base and training data buckets need malware scanning and PII controls.
Good evidence: S3 bucket policies with encryption, access controls, and malware scanning configured.
STD-30
How is the supply chain assessed for security?
Why: Every dependency is a trust decision — unidentified dependencies are unassessed risks.
Good evidence: Dependency inventory with security assessment of critical third parties.

Testing & Sign-off

5 controls

How the system is validated before production, and who signs to say it's ready.

STD-31
What is the testing and evaluation strategy?
Why: Testing proves the system does what it claims — without a strategy, tests are anecdotes.
Good evidence: A testing strategy covering functional, security, and evaluation testing.
STD-32
How are threat model findings validated by test cases?
Why: A threat model without tests is a hypothesis — tests make it evidence.
Good evidence: Test cases mapped to threat model findings with pass/fail results.
STD-33
Who signs off security test results before production?
Why: Someone must be accountable for the test evidence — a named human, not a process.
Good evidence: Named sign-off authority with date and evidence reference.
STD-34
How is the threat model revalidated for production?
Why: Threats evolve between MVP and production — the model must be current at the gate.
Good evidence: Threat model review date showing revalidation before the production gate.
STD-35
How is resilience and recovery validated?
Why: A system that cannot recover from failure is not production-ready.
Good evidence: Recovery testing results, backup verification, and RTO/RPO documentation.

The Gate's Three Asks

3 controls

The three questions every inherited form forgot. These apply universally, regardless of use case size or risk tier.

GATE-1
Are quality acceptance thresholds met?
Why: The gate must know what 'good enough' was measured against — the acceptance criteria and their results, not a feeling.
Good evidence: Named thresholds with measured results for each acceptance criterion.
GATE-2
Is the human-review design confirmed?
Why: Who reviews what, at what confidence threshold, with what override audit — named before go-live, not improvised after.
Good evidence: Named reviewer role, confidence threshold, and override audit log design.
GATE-3
Is a run-owner named and the operate model funded?
Why: A system promoted to production without a named owner and funded operations is an orphan with uptime.
Good evidence: Named owner with budget allocation for ongoing operations.

Depth is derived, not assumed

A low-risk internal tool answers fewer questions than a customer-facing system processing personal information. The classification — not a checklist — decides which additional controls apply. A small organization with a simple use case answers ~25 controls. A regulated enterprise with an agentic system answers ~75. Both are honest. Neither is a shortcut.

Machine-readable

This standard is available as JSON for agents, auditors, and integrations:

GET /api/standard

UseNorth is a product of Lodecrest (Pty) Ltd (registration number 2026/557367/07, South Africa).