ai-safe2-framework

Current State and Release Gap

Reviewed: 2026-10-08

What exists now

Area Current capability Release status
AI SAFE2 CLI Consolidated scanning, gates, evidence, decisions, MCP controls, acceptance and self-check Available; pin 1.0.1
NEXUS Runtime/distributed trust and enforcement option Adapter manifest included; deployment-specific integration remains separate
Skills Agent-facing governance and development methods Supported surface; individual harness installation remains separate
Developer Workflow Provider-neutral risk routing, evidence contracts, PR templates, adoption guide, and generated repository explanation Contract preview available; repository-specific evidence bridge remains to be implemented
MCP Pro Public professional MCP product for coding tools External product surface; adapter contract included
Agency Check Public everyday decision experience and Decision Card model Deployment in progress; adapter contract included
Codex and Claude Both can continue coding and reviewing Optional adapters; no configuration takeover
Sovereign Runtime Runtime integration pattern Adapter placeholder and contract included
Drift/Love Equation Drift and human-alignment evidence source Adapter placeholder and contract included

Gaps before a general-availability release

Must close after 0.1

  1. Implement and test provider execution adapters instead of only manifests.
  2. Add signed or otherwise verifiable evidence integrity beyond content hashes.
  3. Integrate AI SAFE2 CLI-native evidence manifest and replay commands end to end.
  4. Define the production Agency Check API/MCP contract and privacy retention behavior.
  5. Verify MCP Pro entitlement, authentication, rate-limit, and hosted-endpoint behavior against the deployed service.
  6. Add end-to-end canary repositories for game, mobile, API, and AI-agent profiles.
  7. Conduct a security review of workflows, untrusted pull-request handling, secrets, artifact publication, and adapter execution.
  8. Establish upgrade compatibility tests and rollback fixtures.
  9. Add an evidence-backed calibration dataset for provider findings.
  10. Reconcile public AI SAFE2 version, control-count, licensing, and product descriptions before broader marketing.
  11. Add a scheduled, read-only upstream update monitor that reports new pinned dependency versions without applying them automatically.

Adoption experience now delivered

The package now explains what it changes before installation and generates a repository-specific report after configuration. The report shows the active profile, routes, critical paths, review options, authority boundary, preview limitations, and safe extension points. The starter workflow places this report in the GitHub Actions summary and retains it as an artifact.

This closes the discoverability gap. It does not close the evidence-bridge or provider-execution gaps above.

The bundled CI execution receipt closes a narrower attribution gap: adopting repositories can now distinguish environment, transport, harness, assertion, post-processing, and publication outcomes before mapping a deterministic check into an evidence envelope. It does not collect provider results automatically and is not the general-purpose evidence bridge.

0.1 exit criteria

This preview is ready to share when:

Claims boundary

This preview demonstrates structure and executable validation. It does not establish that every listed adapter is implemented, that a project is secure, or that any organization is compliant or certified.