Skip to content

Releases: AdamDiStefanoAI/acr-standard

ACR Control Plane v1.1.6

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 27 Oct 21:46
0db0d35

Software patch release from reviewed merge commit 0db0d354ba44eeb3ca7053f431e442ca6634cec1 (PR #43). This does not revise the normative ACR Standard or claim complete conformance.

Changes

  • Python SDK clients and guards fail closed on executable decisions carried by incompatible HTTP responses, missing explicit MODIFY replacements/parameters, and replacement actions for a different tool. Legitimate unchanged/empty parameters, aliases and structured DENY/ESCALATE remain supported.
  • Upgrade the matching OpenTelemetry family to API/SDK/exporter 1.45.1 and instrumentation 0.66b1, addressing CVE-2026-114603 without weakening the dependency audit.
  • Include previously merged README/newcomer validation and quickstart execution-count fixes, with the limits of the demonstrations stated explicitly.

Validation

Independent exact-diff and merged-tree reviews passed. All candidate and post-merge gates passed, including Python 3.11/3.12, integration/OPA, security, Docker builds, public README and disposable full-stack validation. The fresh base/dev suite recorded 420 passing tests and a clean strict dependency audit. Independent offline observability smoke exercised ACR/HTTPX/FastAPI spans and both OTLP exporter APIs.

Important limits

  • Optional legacy integrations and the pre-1.0 LangGraph family are NOT audit-clean. Do not install them for production evaluation pending the separate migration. Base installation does not install LangChain/LangGraph; this patch does not remediate optional dependencies.
  • SQLAlchemy 2.1 automatic tracing is skipped by the existing instrumentation guard. Complete SQL tracing is not claimed.
  • Response validation does not authenticate a compromised trusted gateway, reverse server-side effects, or solve approval replay/concurrency and durable-evidence failure gating. Enterprise effects in full-stack tests are synthetic. Production deployment still requires its own security review.
  • Signed compliance assets cover their manifest-listed documents, not every repository file or a conformance certification.

Publication verified — 2026-10-27

All five jobs in release run 40589566170 completed successfully.

  • PyPI 1.1.6: clean base-only installation and pip check passed. The tagged no-Docker example verified refund_customer=1, delete_customer=0; this uses synthetic HTTP decisions, not a real policy engine. Another 32 installed-wheel sync/async cases passed, and 94 packaged source files match the reviewed tag.
  • Wheel SHA256: b1dabe5948ff48e07ec7274932af76a24b11a7f7923a7bd95fed82c043e0d2f5; source archive SHA256: 64764e060a17a10b66be1bfaea6f09afa7298eb23b3ef7601ba963a7eec602a3. Both PyPI publishing attestations verified.
  • Gateway image digest: sha256:5e68e9fe6b3861e52fea91e1093c34c63c8dca1cb2449fb94fac33acbfafce6f; killswitch image digest: sha256:4b3a83322fcc8f7a892d09f7a3ef236cd12e45ef82ed87230fba8cc2fd0336c4. Anonymous registry readback and signed SLSA provenance verified the expected workflow, tag and commit.
  • Compliance archive signature and checksums verified; all 12 manifest files match tagged source. Archive SHA256: bbb94c6585dd2a09101ada08d69d87b29efb5b30e5e4f01c68b651566f611958.

Artifact verification is not a production deployment test; the limits above remain.

What's Changed

Full Changelog: v1.1.5...v1.1.6

ACR Control Plane v1.1.5

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 19 Oct 05:55
1247450

Software release from reviewed commit 1247450. This does not change the normative ACR Standard v1.0.1.

Changes:

  • Runnable deterministic LangGraph StateGraph/ToolNode example and sync/async authorization regressions.
  • Structured HTTP 401 responses for missing, invalid, or payload-mismatched protected-executor credentials, with explicit audience/scope/binding/replay limitations.
  • Correct OPA container health probing and Compose forwarding of executor-credential and kill-switch secrets.
  • Disposable full-stack lifecycle validation for all four decisions, approval resolution/rejection/repeats, exact synthetic tool payloads, tampering, outages/recovery, containment, and correlated evidence export.
  • Normative documentation reconciliation, repaired onboarding links, troubleshooting guidance, and an authentic evidence walkthrough.

Validation:

  • Independent bounded source and release-artifact reviews.
  • 353 independently executed regression tests with 81.81% coverage, including optional installed LangGraph dependencies; standard CI does not install those optional dependencies.
  • Fresh CI, security, link checks, and full-stack run: https://github.com/AdamDiStefanoAI/acr-standard/actions/runs/39206513209
  • Final candidate wheel smoke: one allowed effect, zero denied effects with explicitly synthetic HTTP decisions; pip check passed.

Evidence walkthrough: https://github.com/AdamDiStefanoAI/acr-standard/blob/v1.1.5/docs/evidence-walkthrough.md

Scope and limits:
Enterprise effects in full-stack tests are synthetic. No production-readiness or complete-conformance claim; executor replay/concurrency safety, all containment tiers, kill-switch timing SLA, and high-risk durable-log failure gating remain outside this validation. An ordinary Docker newcomer run remains unverified in our local environment. The walkthrough explicitly identifies missing DENY persisted-event/export evidence. Examples and deployment assets are in the tagged repository/source distribution, not the installed wheel. Signed compliance assets cover their manifest-listed documents, not every repository document or test artifact.

Publication outputs are produced by the tag-triggered release workflow. Verify Python package, both container images, and signed assets independently before relying on the complete release.

What's Changed

Full Changelog: v1.1.4...v1.1.5

v1.1.4

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 18 Oct 04:47
3a5c874

Adoption release: the repo now sells the idea in 60 seconds.

  • New README lead: guard_tool first, pip install, no-Docker quickstart. Clone-to-deny in under a minute.
  • New examples: quickstart_no_docker.py (mocked gateway, zero dependencies beyond pip), LangGraph and CrewAI guard examples.
  • Repo renamed to acr-standard; all self-referencing links updated.
  • First outside contribution merged: denial-before-execution test coverage from @yuefdev (fixes #23).
  • 11 good-first-issues open and labeled for new contributors.

Try it: pip install acr-control-plane then run examples/quickstart_no_docker.py. Watch an allowed action run and a denied action get blocked before it executes.

What's Changed

New Contributors

Full Changelog: v1.1.3...v1.1.4

v1.1.3

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 05 Oct 07:34
f66593c

Full Changelog: v1.1.2...v1.1.3

v1.1.2

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 05 Oct 07:27
bd3cb70

Fixes Docker builds under tag-based versioning; PyPI publishing is now independent of Docker.

Install: pip install acr-control-plane

Full Changelog: v1.1.1...v1.1.2

v1.1.1

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 05 Oct 07:21
21f604f

Here's what to paste into the "Describe this release" box — keep it short, because the pipeline auto-generates full release notes when it runs anyway:

First PyPI release: pip install acr-control-plane

Packaging changes only — the version now comes from git tags, and every version tag auto-publishes the package to PyPI.

That's it. Hit Publish release, and the tag fires the pipeline.

v1.1.0

Choose a tag to compare

@github-actions github-actions released this 10 Apr 00:44

What's Changed

Full Changelog: v1.0.1...v1.1.0

ACR Framework v1.0 — Autonomous Control & Resilience

Choose a tag to compare

@AdamDiStefanoAI AdamDiStefanoAI released this 16 Mar 04:11
2f33bea

The ACR Framework (Autonomous Control & Resilience) introduces a runtime governance architecture for managing autonomous AI systems in enterprise environments.

As organizations deploy increasingly autonomous AI agents capable of interacting with enterprise systems, traditional governance approaches focused on development-time controls are no longer sufficient.

ACR defines a runtime governance model that enables organizations to:

• enforce AI governance policies during execution
• monitor AI decision behavior
• detect autonomy drift
• contain abnormal system behavior
• preserve human authority over autonomous systems

The framework includes:

• ACR control pillars
• runtime architecture
• control plane design
• production lifecycle model
• STRIKE threat model integration
• NIST AI RMF mapping
• implementation guidance

The full framework specification is available in this repository.

Author: Adam DiStefano