Skip to content

VALIDATION.md

Latest commit

 

History

History
33 lines (23 loc) · 3.14 KB

VALIDATION.md

File metadata and controls

33 lines (23 loc) · 3.14 KB
Raw

Local validation record — Microsoft Agent Framework v0.1.0

Scope: Functional Windows validation was reported on 2026-10-09 against the versions listed below. This historical record is based on operator-provided PowerShell output, not an independent production deployment or authenticated human-review attestation.

Environment

  • Windows, Python 3.12 virtual environment
  • agent-framework-core==1.21.0
  • loopgrid==0.8.0
  • LoopGrid Core 0.8.1-design-partner at http://127.0.0.1:8000
  • Core ready=true; database=ok; signer=local_ed25519; payload_vault=aes-256-gcm; timestamping=not_configured

Validated by operator

  1. Editable install with real framework and SDK: successful.
  2. pytest -q: 28 passed (with real dependencies installed).
  3. compileall -q src examples tests: passed.
  4. examples/native_core_e2e.py: real Microsoft agent tool dispatch, one simulated $25 refund, two model calls, one actual sandbox invocation; lifecycle evidence_complete, applicable coverage 100%, verifyValid=true.
  5. examples/native_core_policy_e2e.py: all six scenarios passed. Blocked, approval-missing, rejected, wrong arguments and injected event-write failure: 0 actual tool executions. Host-approved: 1 actual sandbox execution with complete verified evidence. Blocked policy and review-required states may refuse persistence of a tool_requested event; this is not evidence of a tool side effect. Incomplete decisions can have <100% coverage while still having valid cryptographic chains.
  6. examples/native_core_microsoft_approval_e2e.py: actual Microsoft @tool(approval_mode="always_require") and ToolApprovalMiddleware paused, with 0 executions until resume in the same session. Rejection: 0 executed and signed rejected state with 100% coverage. Approval: 1 executed, host-recorded sandbox outcome, evidence_complete, 100% coverage, verifyValid=true.

All simulated transactions explicitly reported realMoneyMoved=false.

What these validations do NOT prove

  • No authenticated production human reviewer was tested. The host-generated approval responses are synthetic (authenticatedHumanIdentityValidated=false). Production applications must authenticate the reviewer, validate the pending tool name and exact arguments, correlate the request/decision, record the review only after authentication, and prevent replay or approval substitution.
  • Neither the SDK nor middleware executes real payments. Core verification proves integrity of recorded evidence, not the actual correctness of a policy, reviewer identity, or downstream business claim.
  • Streaming response finalization, durability of in-process action budgets across crashes, distributed authorization and production service availability were not validated.
  • The local runtime test record does not itself establish CI status, installation from a public package index, production identity provenance, or compatibility with other LoopGrid integrations.

Reproducing the validation

See WINDOWS-VALIDATION.md for the exact checks. Review GitHub Actions and packaging logs for the version you install; this record only describes the local functional tests above.