ai-safe2-framework

Assessment Scope

safe2 evidence scope creates a metadata-only inventory of the paths declared inside or outside a deployable subject. It prevents a mixed repository from silently treating tests, adversarial fixtures, examples, research, generated artifacts, dependencies, and third-party code as identical production behavior.

This is the first engineering unit planned for CLI 0.6.0. Its merge does not by itself create a CLI 0.6.0 release.

Technology contribution scope

For an artifact review, reuse system identity and scope in the Technology Contribution Profile. Record covered paths/planes, exclusions, external dependencies, and unknowns per control. An included implementation is not a tested mechanism, and an excluded benchmark may supply evidence without becoming deployed enforcement.

Quick start

Create the system identity first:

safe2 evidence system safe2/data/system-identity-source-demo.json \
  --output system-identity.json \
  --strict

Build the scope inventory:

safe2 evidence scope safe2/data/assessment-scope-source-demo.json \
  --project-root . \
  --system-identity system-identity.json \
  --output assessment-scope.json

The packaged declaration is intentionally incomplete for the full AI SAFE² repository. Unmatched paths remain unknown and partial. Add reviewed rules for the actual deployment subject before using --strict.

Classifications

Classification Intended meaning
product Code or configuration shipped as part of the deployable subject
test Test implementation or evaluation support outside runtime behavior
adversarial_fixture Deliberately hostile content used to test a defense
example Reference or instructional implementation
research Research material not declared as deployed behavior
generated Build, cache, report, or local evidence output
dependency Installed or vendored dependency governed as a dependency
third_party External implementation retained with third-party provenance
configuration Configuration whose deployment relationship requires an explicit disposition
documentation Human or agent documentation outside executable behavior
unknown No reviewed rule establishes the path’s deployment relationship

Dispositions

Excluded and not-applicable are not synonyms. An excluded test may still supply evidence about included product code. A not-applicable path requires an explicit reviewed rule rather than absence of a match.

Rule behavior

Rules use normalized, relative POSIX glob patterns and are evaluated in declared order. Absolute paths, drive-qualified paths, backslashes, alternate-stream syntax, and parent traversal are rejected.

When multiple rules match with different classification, disposition, or component bindings, the conflict remains in the artifact. The first rule supplies inventory fields only so the record stays deterministic. Any conflict blocks --strict success.

The default must classify unmatched paths as unknown, with either partial or excluded disposition. Strict mode fails when the result contains unknown, partial, conflicting, truncated, or unsafe-link coverage, or contains no included paths. The JSON artifact is written before the strict exit so the reason remains inspectable.

Security and evidence boundaries

The collector:

Every result states:

{
  "decision_scope": "assessment_scope_inventory_only",
  "scope_verified": false,
  "content_inspected": false,
  "conformance_claim": false
}

Path rules remain attributed declarations. They do not prove build inclusion, runtime loading, execution, authorization, control effectiveness, or AI SAFE² conformance. A digest binds submitted bytes but does not authenticate the declaration’s author.

Contracts

Retrieve them with:

safe2 schema export assessment-scope-source-v1
safe2 schema export assessment-scope-manifest-v1