Assurance cases for AI systems
Claims about a system’s safety and performance for a use are broken into sub-claims, each supported by evaluation evidence, with gaps and assumptions recorded.
GCRI · Artificial Intelligence and Autonomous Systems
AI evaluation and assurance
AI evaluation and assurance engineering is the second method this platform stewards. It provides the evidence that an AI system is fit for a stated use: evaluations, red teaming, documentation, monitoring and the structured arguments that tie them together.
01Register
AI evaluation and assurance on the record.
Method recordSTEM method
- MethodAI evaluation and assurance engineering
- StewardArtificial Intelligence and Autonomous Systems
- Applied inArtificial Intelligence and Autonomous Systems
- Data4 inputs
- Checks3 validation steps
- Canonical pageSTEM Commons
02Application
What AI evaluation and assurance engineering does here.
AI evaluation and assurance engineering develops methods to test, monitor and provide justified confidence in AI systems. It addresses performance, robustness, safety and societal impact across the system life cycle. In artificial intelligence and data and compute, AI assurance determines whether automated systems can be trusted in high-consequence settings. It helps manage risks when AI is embedded across finance, health and infrastructure. The canonical description is on the STEM Commons.
Structured red teaming
Adversarial testing is planned around threat models, with testers, scope, time and findings documented so that coverage can be judged.
Post-deployment monitoring
Incidents, complaints and performance metrics are tracked after release, and triggers for review or rollback are defined in advance.
03Monitor
Evaluation suite.
The screen a team would use to follow this method: its measures, states and decisions. Names are examples and values are illustrative.
Evaluation suite
AI and autonomous systems · AI evaluation and assurance · Run R-0951
One gate pendingGCRI · Artificial Intelligence and Autonomous Systems
Evaluation suites
| Capability | 585 | 89 | Release | Pass |
| Robustness | 133 | 72 | Release | Pass |
| Bias and fairness | 169 | 81 | Release | Pass |
| Misuse red team | 142 | 79 | Deploy | Open findings |
Method card
- Method
- AI evaluation and assurance engineering
- Model
- M-196 v1
- Owner
- model risk function
- Label
- MODELLED
Findings
- F-1Unsafe completion on edge promptsMitigation before deploy5 d
- F-2Lower accuracy on minority dialectData augmentation9 d
04Data
What the method consumes.
Datasets and records used by AI evaluation and assurance engineering on this platform, with the role each plays and the label its values carry. Sources are listed in the source register.
| Data | Role | Label |
|---|---|---|
| Evaluation results | Evidence for claims | OBSERVED |
| Red-team reports | Failure discovery | REPORTED |
| Incident logs | Field evidence | REPORTED |
| System documentation | Scope and assumptions | DECLARED |
05Methods and evidence
How AI evaluation and assurance is modelled, measured and tested.
The model, its assumptions, its data and its uncertainty, in the form a scientist or engineer would check. Structures follow published methods. The values drawn are illustrative, not findings.
Model, equations and assumptions
The model structure used for AI evaluation and assurance, the assumptions it makes and how each is tested.
Risk of element i
Ri = f(Hi, Ei, Vi, Ci)hazard, exposure, vulnerability and coping capacityExpected annual loss
EAL = ∑k pk · Lkover events k with annual probability pkCoupled systems
xt+1 = A xt + B ut + εtA carries the couplings, u the shock, ε the error| Assumption | Why it is made | How it is tested |
|---|---|---|
| Hazard and exposure are independent within an event | Keeps the loss integral tractable | Compare with a copula fit on joint records |
| Couplings are linear near the operating point | Lets shocks propagate through A | Perturb beyond the linear range and compare |
| Vulnerability curves transfer between places | Local curves are rarely available | Validate against observed losses where they exist |
- Inputs
- versioned datasets with provenance
- Outputs
- distributions, never single numbers
- Code
- released with the model card and tests
- Reproduction
- same data, code, environment and seed
FAIR guiding principles · GO FAIR and FORCE11W3C PROV · World Wide Web ConsortiumGuidance on consistent treatment of uncertainties · IPCCCEOS Analysis Ready Data · Committee on Earth Observation Satellites
06Validation
How results are checked.
Evaluation methods are checked for reliability and relevance to deployment conditions. Independent replication and third-party audits test claims.
| Check | How it is done |
|---|---|
| Independent review | Assurance cases are reviewed by people not involved in development |
| Evidence currency | Evidence is re-checked after model or deployment changes |
| Coverage reporting | Evaluations state what they did not test |
Benchmarks may not reflect real-world use or novel capabilities. General-purpose systems have open-ended behaviour that resists complete testing.
07Process
How the process runs.
Process
AI evaluation and assurance: AI risk management (NIST)
The AI RMF Core has four functions, GOVERN, MAP, MEASURE and MANAGE, with GOVERN cross-cutting; the order drawn is one common reading, not a checklist.
01 Govern
- 01GOVERN: policies, roles and accountabilityStep · AI risk committeeGOVERN is a cross-cutting function. It is infused throughout AI risk management and enables the other functions.
02 Map
- 02MAP: context and intended useDatasetThe MAP function establishes the context to frame risks related to an AI system, including intended purposes and impacts on people.
03 Measure
- 03MEASURE: evaluate trustworthy characteristicsModelAI systems are evaluated for trustworthy characteristics. They should be tested before deployment and regularly in operation.
- 04MEASURE: track identified risks over timeModelMechanisms for tracking identified AI risks over time are in place, including unanticipated and emergent risks.
04 Manage
- 05MANAGE: treat risksStepRisk treatment plans cover response to, recovery from and communication about incidents. Resources go to mapped and measured risks.
- 06Proceed or not (MANAGE 1.1)Gate · AI risk committeeA determination is made whether the system meets its purposes and objectives, and whether to proceed. The framework sets no thresholds.
- 07Deployment decisionDecision · AI risk committeeThe organisation that owns the system decides whether and under what conditions to deploy.
- 08Post-deployment monitoring and incidentsModelPost-deployment monitoring plans are implemented. Incidents and errors are shared with relevant AI actors, including affected communities.
05 Record
- 09Inventory of AI systemsDatasetGOVERN 1.6 asks for mechanisms to inventory AI systems. A model card is common practice and is not named in the framework.
Neutral outline of a published external standard or method, from its issuing body. Not a record of operation. Nexus does not perform, certify or approve any step shown. A gate or decision belongs to the party the source names. The actions are not necessarily an ordered set of steps. The EU AI Act is not drawn in this flow.
08Connections
Topics that rely on this method.
09Stewardship
Who maintains the method.
Artificial Intelligence and Autonomous Systems stewards AI evaluation and assurance engineering for the Consortium and reviews how it is applied across platforms.
10Questions
Questions and answers.
What are the limits of AI evaluation and assurance engineering in this field?
Benchmarks may not reflect real-world use or novel capabilities. General-purpose systems have open-ended behaviour that resists complete testing.
How are results checked?
Evaluation methods are checked for reliability and relevance to deployment conditions. Independent replication and third-party audits test claims.
Are the values on this page real?
The data names, roles and checks are real descriptions of practice. The canvas shows a design of the method’s use, not a run; no result is reported here.
Where else is this method applied?
The canonical method page lists every platform that applies it.
ReportsContribute
Contribute evidence from this page.
Reports is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.
GCRIArtificial Intelligence and Autonomous SystemsExponential TechnologyLabs Evidence
- EmailGateway GW-GCRISend it from this pageEvidence, a correction, a case, a method, a review, a translation or profile input. It arrives tagged with this page, so the editor who owns it receives it.
- ZenodoOpen depositDeposit it and submit to the communityPublish your own record on Zenodo with your ORCID iD, then submit it to the Reports community there. Curators review it; the DOI stays yours.
- ReportsGCRISee every way to contributeThe contribution types, how each is reviewed, what credit you receive, the licence and what is never accepted.
Never send personal data, sensitive infrastructure detail or anything you have no right to share. A contribution is reviewed before use; sending it confers no authorship or editorial rights, and sponsorship never buys a finding.
Artificial Intelligence and Autonomous Systems
Bring a method, a dataset or a question.
Information and enquiries are open today. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.
Legal
Legal documents.
Maps, data and figures on this site are shown for context and general information. They are not advice, forecasts or warnings; the limits of the information are set out in the Disclaimers and Data Notices.