AI evaluation and safety evidence
Evaluation plan, benchmark, run, results, risk assessment and decision on permissible use. Metric, dataset and evidence are reused.
Bundle → Layer → Finding → Questions Filled
2 bundles · 4 layers · 5 findings · 12 questions
Evaluation What was measured, on what, and under which plan.
Plan and benchmark
The evaluation plan and the benchmark specification it uses.
Evaluation plan
Purpose, context, metrics and thresholds fixed before the run.
- Which context of use and risks does the plan address?
- Which metrics and thresholds were fixed before the run?
Benchmark integrity
The benchmark version and its separation from training data.
- Which benchmark and dataset versions were used?
- Has the benchmark leaked into the training data of the model?
Run and results
A concrete evaluation run with pinned versions and uncertainty.
Evaluation run
Results tied to model and dataset versions with stated uncertainty.
- Which model and dataset versions does the result pin?
- What uncertainty or variance is reported for the result?
- Can the run be repeated with the same versions?
Assessment and decision How results become a decision about permissible use.
Safety assessment
Risk assessment in a stated context, separate from capability scores.
Context-bound assessment
Risks judged for a specific use, users and setting.
- For which application context is the assessment valid?
- Which risks were assessed beyond the measured capability?
Deployment decision
The decision on permissible use, with owner and conditions.
Decision record
Who decided, on what evidence, under which conditions.
- Who is the accountable owner of the decision, and on which evidence did they rely?
- Which conditions and limits does the approval carry?
- Does a change of model, dataset or context require a new evaluation?
Classifiers Filled
- Family
- Enterprise profiles
- Category
- Enterprise subject
- Entry kind
- subject
- Domain
- EnterpriseAI
- Industry
- Cross-industry
- Tags
- EM-AI-03W3subjectEvaluationPlanEvaluationRunBenchmarkSpecificationSafetyAssessmentDeploymentDecision
What it is Filled
Evaluation plan, benchmark, run, results, risk assessment and decision on permissible use. Metric, dataset and evidence are reused.
Why it exists Filled
Evaluation plan, benchmark, run, results, risk assessment and decision on permissible use. Metric, dataset and evidence are reused.
Distinguishing features Filled
- An evaluation result pins the model and dataset versions it was obtained with.
- Capability measurement and permission to use are separate records with separate owners.
- An evaluation holds only within its stated context and does not transfer to other uses.
- Distinct from the model artifact record, which describes what was trained rather than how it performed.
What robots and AI may and may not do Filled
Must not
- Declare a system safe for all tasks on the basis of one high benchmark score.
- Carry an earlier approval over to a changed model, dataset or context of use.
- Record a deployment decision without an accountable owner and its conditions.
- Report results while omitting uncertainty or failed runs.
Only with a human decision
- Approving a model for a new context of use.
- Accepting residual risks identified in a safety assessment.
May
- Record evaluation runs with pinned model and dataset versions and reported uncertainty.
- Check benchmarks for overlap with training data and report suspected leakage.
- Flag a deployment whose model, data or context no longer matches its approval.
Moral aspects Filled
- Overstated evaluations can expose people to unsafe automated decisions.
- Accountability requires that a named person owns each decision to deploy.
Who is affected
- People affected by the deployed system
- Operators relying on the approval
- Evaluators and reviewers
Owners Filled
Steward
The head of AI or research who answers for models and their evidence on behalf of the organisation.
Master systems
- Model registry
- Experiment tracker
- Evaluation store
Links to other meta-models Filled
neighbor
- EM-AI-01
- EM-AI-02
references
- WM-AI-003 - conceptual-candidate
- WM-AI-009 - conceptual-candidate
- WM-AI-008 - conceptual-candidate
- WM-AI-010 - conceptual-candidate
What else AI and robots need to interact with it Filled
Identity and identifiers required Filled
- Identified by evaluation run identifier together with the model and dataset versions it pins.
Direct properties not applicable Not applicable
Not applicable
Enterprise record contour: physical properties belong to referenced world models.
Recognition optional Filled
- An evaluation record states a plan, a context, versions and results with uncertainty.
- Often confused with a published leaderboard score or with the deployment decision itself.
Capabilities and actions required Filled
- Results can be compared across model versions on the same benchmark version.
- Decisions can be audited back to the evidence they relied on.
Hazards and failure modes required Filled
- Benchmark contamination that inflates scores.
- Approval drift when context or versions change without re-evaluation.
- Selective reporting of favourable runs.
Standards and interfaces required Filled
- NIST AI Risk Management Framework for evaluation and risk management.
- ISO/IEC 42001 for AI management systems.
- W3C PROV-O for provenance of evaluation evidence.
Context of use required Filled
- Used before and after deploying AI systems, and whenever their context of use changes.
- High-risk uses require conformity evidence under the EU AI Act.
Sources Filled
- NIST AI RMF: context, evaluation and risk management
- PROV/SPDX: provenance of data, artifacts and actions
- Model registry/evaluation tooling practice; map system, artifact, run and endpoint
- NIST AI Risk Management Framework - Checked live (HTTP 200) on 2026-10-07.
- NIST AI RMF Playbook - Checked live (HTTP 200) on 2026-10-07.
- W3C PROV-O: The PROV Ontology - Checked live (HTTP 200) on 2026-10-07.
- SPDX Specification - Checked live (HTTP 200) on 2026-10-07.
Open questions
- How to detect benchmark leakage into training?
- How to separate capability measurement from permissibility of use?
- How to account for uncertainty and a context change?
- Установить границу и решение reuse/extend/new по действующим спецификациям.
- Подтвердить semantic crosswalk, права и source mastership.
- Выбрать immutable refs; провести проверки fixtures до заявления о публикационной готовности.
Machine files
Provenance
enterprise research programme · published-partial
Built from: enterprise/models/em-ai-03/brief.json, ver-cy/world-models/card-supplements/em-ai-03.json
Still in Russian: suggested owner, blocking decisions, vercy candidates. Translate in research/enterprise/i18n/units.en.json.