| Framework Home | Cross-Pillar Governance | AISM | NEXUS | Technology Profile |
Method version: 1.0
Documentation updated: 2026-10-13
Status: CSI assessment method for use and feedback
Start with the AISM assessment guide for a practical workflow. Use the technology evidence request to collect versioned artifacts from a provider or internal owner before rating claims.
This v1.0 evidence rubric is separate from organizational maturity, contribution capability, framework conformance, and Challenge Lab claim maturity. Always name it AI SAFE² Evidence Assurance v1.0 with each E rating.
| Rating | Evidence class | Minimum record |
|---|---|---|
| E0 | Assertion | Attributed marketing or unsupported description; no supporting artifact |
| E1 | Design | Identifiable architecture, paper, or control specification supporting the design claim |
| E2 | Implemented | Inspectable implementation/released code with version and mechanism traced to the claim |
| E3 | Controlled test | Defined harness, positive and negative cases, environment, reviewable results, and explicit outcome/coverage |
| E4 | Independent validation | Named independent evaluator and reproduced/evaluated material claim, artifacts, method, and limitations |
| E5 | Operational evidence | Representative production evidence over a stated period, failures, incident handling, and coverage gaps |
Record the highest class substantiated for the specific claim. These classes are not a universal quality ordering: operational evidence may be self-reported, and a controlled test can be narrow. E5 does not imply E4 independence. Preserve independence as a separate field. A paper can contain E3 evidence only when its test method and material results are reviewable; publication alone supports E1. Evidence of a failure can qualify as E3 or E5 without validating prevention.
Unknown or unreviewed evidence has not_assessed, no rating, and a reason. E0 is
an observed unsupported assertion, not a substitute for an unperformed review.
Different claims about the same technology may receive different ratings.
For each claim retain:
Preserve superseded evidence with its status. Do not upgrade a rating because a source is popular, open source, signed, translated into SAFE2, or vendor-published. A digest establishes byte consistency; it does not establish truthful observation. Generalizing from one environment requires additional evidence.
The current AISM CLI model maps E0-E5 to numeric
weights. scoring.py combines those weights with
verification caps and evidence-category completeness. Those input and summary
grades are not automatically assurance-v1.0 ratings. This method changes no
weights, caps, formulas, schemas, or previously issued results.
Store assurance assessments in the separate Technology Card.
Do not copy their ratings into existing CLI grade fields without a separately
reviewed translation. A derived CLI evidence-summary grade must not be relabeled
as an independently assessed assurance level.
Challenge Lab C0-C5 describes progression of a control claim through implementation, scenario testing, bypass testing, and replication. Record the applicable challenge result alongside assurance, preserving its own criteria. Neither C5 nor E4 establishes an organization’s AISM Level 5.
| Technology Profile | Validation | Examples |
AI SAFE² v3.1 · Cyber Strategy Institute