AI Certification Authority

The standard for
AI in critical
systems

Every standard currently applied to AI in human spaceflight was written before AI existed in human spaceflight. Safety Critical Labs provides the first certification framework built specifically for AI in safety-critical operations.

View the framework Read the standard →
Assessment — System 07-A In review
AI-1 Data partitioning Pass
AI-2 Bias detection Pass
AI-3 ML test coverage Pass
AI-4 Continuous validation Review
AI-5 Hallucination prevention Pending
AI-6 – AI-9 Remaining areas Queued
CL-ASSESS-2025-001 Human Spaceflight / Safety-Critical
The gap

Existing standards
weren't built for AI

NPR 7150.2D, ECSS-E-ST-40C, and ISO/IEC 5338 were written for deterministic systems. None address the failure modes that make AI in critical systems genuinely dangerous.

NPR 7150.2D
V&V occurs at delivery
Assumes stable behavior post-deployment. AI systems drift. No mechanism exists for re-validating a degraded model mid-mission.
All spaceflight standards
Hallucination is unrecognized
Confident, plausible, incorrect outputs are not a recognized failure mode in any current spaceflight software standard.
NASA-STD-8739.8
Safety cases assume code inspection
Safety analysis requires logical traceability. Black-box AI outputs cannot satisfy this requirement without an explicit explainability layer — none is mandated.
ISO/IEC 5338 & 24027
Guidance without criteria
The closest AI-specific standards are process guidance documents. They name concerns without providing verifiable pass/fail criteria.
All standards
OOD detection absent
No standard requires detection of inputs outside a model's training distribution — the condition under which AI behavior becomes least predictable.
Human spaceflight
No AI-specific certification path
There is no established process by which an AI system in a human-rated vehicle can be assessed, certified, or recertified against a defined standard.
The framework

Nine requirement areas.
Every AI failure mode covered.

The Safety Critical Labs framework establishes verifiable requirements for each AI-specific failure mode not addressed by existing spaceflight standards. Algorithm-agnostic and domain-applicable.

AI-1
Data partitioning & classification
Training/validation/test separation with documented provenance. CUI and ITAR classification inheritance through AI output channels.
All classifications
AI-2
Bias detection & mitigation
Baseline performance across user classes and operational contexts. Bias-free training data requirements with verifiable inspection criteria.
All classifications
AI-3
ML test coverage
A defined test matrix as the ML equivalent of code coverage — spanning nominal performance, edge cases, and failure mode injection.
All classifications
AI-4
Continuous validation
Post-deployment drift detection with defined thresholds, revalidation triggers, and escalation paths. Closes the single largest gap in existing spaceflight standards.
All classifications
AI-5
Hallucination prevention
Output confidence bounding with defined thresholds for safety-critical decisions. Graceful degradation to human override below threshold.
All classifications
AI-6
Out-of-distribution detection
Detection, logging, and escalation of inputs outside the training distribution — the condition under which AI behavior is least predictable.
All classifications
AI-7
Adversarial robustness
Adversarial testing, defined robustness bounds, and specified behavior under adversarial input conditions for safety-critical deployments.
Safety-critical only
AI-8
Explainability
Operator-accessible decision reasoning with traceable decision basis. Required for any AI system where a traditional safety case would require logical inspection.
Safety & mission critical
AI-9
Human-AI teaming
Defined AI authority limits, human override requirements, trust calibration testing, and interaction audit trails. Built for spaceflight crew dynamics.
Safety & mission critical
Certification process

From assessment
to certification

A rigorous four-phase process produces a formal determination — not a recommendation — that your system was assessed against a defined standard.

01
Applicability determination
Classify your AI system against the framework criteria. Establish which requirement sets apply and document any tailoring rationale.
02
Evidence package review
Submit documentation, test reports, and verification matrices. Our team reviews against each applicable AI-1 through AI-9 requirement.
03
Technical assessment
Structured audit of your system against the framework. Findings documented with specific requirement references and success criteria.
04
Certification issuance
Formal determination document stating which requirements were met, at which classification level, and as of which framework version.
Our principles

How we work

Certification is only meaningful if the standard behind it is rigorous, transparent, and built for the domain it covers.

◆
Open standard The Safety Critical Labs framework is publicly available on GitHub. We certify against a standard anyone can read, audit, and challenge. Transparency is not optional — it is the foundation of trust.
◆
Determinations, not recommendations A Safety Critical Labs certificate states what was assessed, against which version of the framework, and at which classification level. It does not claim a system is safe — it claims the system was rigorously evaluated against a defined standard.
◆
Domain-first methodology The framework was developed from within the operational environment of human spaceflight — not adapted from a generic AI governance document. Requirements address the actual failure modes of actual systems in actual use.
◆
Certification is not a one-time event AI systems drift. A certificate issued at deployment is not a certificate valid at month six. Safety Critical Labs builds continuous validation requirements into every assessment — because the standard demands it.
Start here

Ready to certify
your AI system?

Whether you're preparing for regulatory scrutiny or building responsible AI practices into your development lifecycle now — we can help.

Get in touch Read the framework →