ASSERT v0.3.0 Release

ASSERT v0.3.0

ASSERT v0.3.0 improves dependency isolation, evaluation authoring, sandbox security, and result exploration. This release includes breaking changes to optional dependency names and the default multi-turn inference budget.

Highlights

• Evidence-backed eval config generation in the run-assert-eval skill, including source research, citation review, and one-risk-per-suite output.
• Host-authoritative sandbox action mediation, with case-correlated decisions and host-owned evidence.
• Atomic behavior presets with broader built-in risk coverage and CI enforcement.
• Cleaner dependency ownership between ASSERT capabilities and target frameworks.
• Improved viewer tables and metric dropdown behavior on constrained screens.

Breaking changes

Optional dependencies

ASSERT extras now represent product capabilities instead of repository examples:

Previous	Replacement
assert-ai[otel]	assert-ai[phoenix] plus the target framework's OpenInference instrumentor
assert-ai[azure-aad]	assert-ai[azure-auth]
assert-ai[langgraph]	Install target-owned dependencies
assert-ai[dspy]	Install target-owned dependencies
assert-ai[examples]	Use each example's requirements.txt
assert-ai[regression]	Use the repository development dependency group

The all extra now installs all ASSERT capabilities: Phoenix, analysis, embeddings, Azure authentication, and ACS integration.

Multi-turn inference

The default pipeline.inference.max_turns value changes from 10 to 6.

Set the value explicitly to retain previous behavior:

pipeline:
  inference:
    max_turns: 10

Results generated before and after this change may not be directly comparable for multi-turn behaviors.

Security and reliability

• Sandbox tool decisions can be evaluated and recorded by an authenticated host service.
• Network evidence is stored outside the evaluated target's writable mount.
• Sandbox mediation now correlates policy, mocks, and evidence by test case.
• Real Docker containment tests run in CI.
• LiteLLM and frontend dependencies were updated to patched versions.
• The HTTP endpoint target now declares its direct aiohttp dependency.

Additional improvements

• assert-ai --version now reports the installed package version.
• Traced callable targets emit an actionable warning when no spans are captured.
• Analysis, Phoenix, embeddings, and framework dependencies are isolated into compatible environments.
• Documented example environments receive clean-install and pip check coverage.

Upgrade

python -m pip install --upgrade "assert-ai==0.3.0"

Review the breaking changes above before upgrading existing evaluation environments.