ai-safe2-framework

AISM Technology Contribution Profile

Assess a technology’s contribution without assigning it organizational maturity

AI SAFE² Method Status

Framework Home Cross-Pillar Governance AISM NEXUS Research Challenge Lab

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.

Purpose and status

An organization earns an AISM maturity level. A product, paper, benchmark, investigation, or technique contributes evidence toward specific controls. The organizational model’s full name is AI Sovereign Maturity (AISM) Model. This v1.0 review instrument, AISM-TCP, separates that contribution from organization-wide scoring. CSI introduces it as a method for use and feedback. It does not confer certification or add a CLI feature or framework control.

The primary reader is the security leader deciding which control gap to address. The new decision is whether a candidate supplies the needed mechanism and evidence, not whether its vendor has the highest aggregate score. This extends existing runtime-control and research guidance with a scoped adoption method.

Instrument Subject Result Does not establish
AISM Sovereignty Score Organization across five pillars and six dimensions Maturity using coverage, robustness, and sovereignty assurance Product certification or framework conformance
AISM Technology Contribution Profile, v1.0 Exact artifact and deployment/test boundary Control effects, pillar contribution, coverage, limitations Vendor or adopting organization maturity
AI SAFE² Evidence Assurance Level, v1.0 Each attributed claim and supporting artifacts E0 through E5 under the named assurance rubric Control completeness, independence by default, or universal prevention

Report all three independently. Do not multiply contribution by assurance, average pillars to rank purchases, or substitute the maximum pillar contribution for an organizational level. Existing assessment formulas, 161 core controls, profile counts, and component versions remain unchanged.

Assessment workflow

  1. Identify artifact type, owner, exact version/commit, assessment date, assessor, source artifacts, and whether evidence is self-reported or independently reviewed.
  2. Define the complete agent system and assessment scope: model, harness, tools, skills, memory, policy, credentials, OS/kernel, infrastructure, and external services.
  3. Select actual required controls from the core dataset and applicable profiles. Record canonical ID, name, document, and version. Declare north-south, east-west, and agent-to-tool scope separately.
  4. Classify the effect of each claim. Link artifacts and the assurance rating to that claim, not to the artifact’s reputation or the vendor’s whole portfolio.
  5. Record contribution, control coverage, failure behavior, bypasses, and gaps. Separate demonstrated, author-reported, intended, contradicted, and unknown facts.
  6. Choose an adoption decision with a named owner, complementary controls, acceptance tests, residual risk, and a review trigger.
  7. Publish a Technology Card and preserve the source/version trail. Reassess after changes to any component or dependency.

Contribution rubric

Each pillar entry contains a score and status, specific mapped claims, coverage, evidence references, and reasoning. Scores describe the highest demonstrated capability within that stated boundary; they do not imply pillar-wide coverage.

Score Contribution Required demonstration
0 None Reviewed artifact has no material relationship to this pillar
1 Conceptual Identifies a risk, principle, or proposed approach
2 Visibility Produces inventory, telemetry, analysis, or threat evidence
3 Governance Supplies repeatable policy, procedure, assessment, or validation
4 Control Enforces or verifies a runtime control with measured outcomes
5 Sovereignty Demonstrates continuous adaptation, formal verification of critical behavior within scope, and durable evidence while retaining organizational authority

The names parallel the maturity ladder, but contribution 1 means Conceptual, not organizational Level 1 Chaos. Use contribution 4, not AISM Level 4, for technology results.

Use not_assessed with no numeric score when evidence is missing. Use not_applicable with a reviewed reason when the pillar is outside the review’s purpose. Neither state is 0. Scores 4 and 5 require reviewable measured results under assurance v1.0, at least E3 for the claimed behavior; E4 and E5 cannot be inferred from the contribution score. A paper’s proposed runtime architecture alone supports a design claim, not demonstrated contribution 4.

For contribution 5, provide repeated challenge/update history, the formal property and its proof boundary, protected evidence retention, named change/stopping authority, and inspect/configure/revoke/replace/export capability. Continuous benchmarking alone or an adaptive model alone is insufficient. A verifier may support contribution 4 without preventing an action itself; its effect remains validates, and an enforcement partner must be identified.

Control effects and coverage

Effect Meaning Required distinction
Addresses Relevant to the control Relevance is not implementation
Implements Contains a functioning mechanism Source availability alone does not prove operation
Enforces Prevents or terminates the prohibited effect within scope Identify interception point, decision mechanism, actuator, and bypass paths
Validates Tests another control’s operation A test is not the enforcement mechanism
Evidences Produces records of operation or failure Logging does not establish prevention
Challenges Shows an assumption or control is insufficient A finding is not a containment product

Multiple effects are allowed when each is supported separately. Coverage is full, partial, none, unknown, or not_applicable for the named control and boundary, with an explicit denominator where quantified. One enforced path does not cover other tools, kernels, cloud APIs, descendants, delegated agents, or planes. Keep required-control applicability distinct from artifact relevance.

Record deterministic interception, probabilistic authorization judgment, and deterministic actuation separately. A model-based ALLOW/DENY decision does not become deterministic merely because its result is enforced outside the agent. Identify failure on timeout, malformed output, policy unavailability, revocation, restart, and enforcement outage. Do not infer fail-closed behavior from normal tests.

Validation and incident evidence

Apply the validation guidance before claiming measured runtime control. Use clean/attacked pairs, independent observation of side effects, direct and composed bypasses, outage cases, and legitimate-task utility measures. Keep Challenge Lab C0-C5 distinct from assurance E0-E5 and contribution 0-5. There is no automatic conversion among them.

Incident sources may challenge a control or drive governance improvements. Preserve activity reconstruction, affected-party identification, retention/hold decisions, notification ownership, and regression tests through the incident evidence workflow. A notification or investigative subpoena does not prove a breach, misconduct, or control effectiveness.

Purchasing and adoption rule

Evaluate candidates in this order, preserving unknowns as decision conditions:

  1. Required-control fit: Which actual gap is covered, and how much?
  2. Enforcement location: Is the boundary outside the governed component’s authority?
  3. Evidence strength: Which named claims are designed, implemented, tested, independently reproduced, or observed in representative operation?
  4. Failure behavior: Does the mechanism fail closed, fail safe, or disappear?
  5. Sovereignty: Can the organization inspect, configure, revoke, replace, and export evidence without vendor permission?
  6. Residual risk: Which attack paths and dependencies remain uncovered?
  7. Integration value: Which other pillars improve, and which new dependencies appear?

A narrow point solution may be the right choice. An unknown prerequisite or critical uncovered path can block adoption regardless of its maximum contribution. Record the resulting decision as adopt, pilot, hold, or reject, with conditions, decision owner, review date, and withdrawal/rollback plan.

Permitted claims and examples

Permitted after supporting tests: “This version supplies contribution 4 evidence for the named Shield and Circuit Breaker controls within the tested Linux boundary.” Include the assurance rubric, assessor, test artifacts, date, coverage, and limits.

Prohibited: “ContractWarden is AISM Level 4,” “DGF-Bench proves Level 5 governance,” “open source means independently validated,” or “a passing vendor test establishes our organization’s maturity.”

The nine-example review and adoption backlog uses the supplied recommendations to frame acceptance work. Those examples are source-attributed candidates, not scored certifications or completed integrations.

Compatibility and implementation boundary

The CLI already uses E0 through E5 as numeric weights in aism-model-v1.json. Its evidence summary also applies verification caps and category completeness. Assurance v1.0 is a separate claim-level rubric; it does not redefine those weights or the existing grade field. Do not put a Technology Card or its contribution into an assessment cell.

Existing safe2 aism ingest imports attributed evidence with cells unscored for review. Assessors still supply organization-specific coverage, robustness, and sovereignty assurance under the current methodology. A future TCP schema, command, adapter, dashboard view, or grade translation requires a separate versioned contract, explicit migration policy, and tests. This update delivers the review method and documentation, not those runtime features.


Technology Card Evidence Assurance Validation Examples

AI SAFE² v3.1 · Cyber Strategy Institute