chore(release): prepare v0.3.0

.github/workflows/science.yml

@@ -54,7 +54,7 @@ permissions:
 env:
   AZURE_API_KEY: ${{ secrets.AZURE_API_KEY }}
   AZURE_API_BASE: ${{ secrets.AZURE_API_BASE }}
-  AZURE_API_VERSION: "2024-12-01-preview"
+  AZURE_API_VERSION: "2025-03-01-preview"
   # Long-context judge required — gpt-5.4-nano/mini fail mechanically on
   # realistic agent transcripts (>10 turns) by not returning the required
   # verdict schema. See user note 2026-05-09.

CHANGELOG.md

@@ -9,11 +9,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
-- Clean-install checks for every documented ASSERT/example environment, including `pip check` and credential-free target imports.
+### Changed
+
+### Fixed
+
+## [0.3.0] - 2026-09-04
+
+Adds evidence-backed eval authoring, stronger host-authoritative sandbox mediation, a broader atomic behavior library, and cleaner optional dependency ownership. This release also changes the default multi-turn budget and includes security, compatibility, and viewer fixes.
+
+### Added
+
+- Evidence-backed eval config generation for the `run-assert-eval` skill, including primary-source research, citation review, human approval before write, and one-risk-per-suite output (#337).
+- Atomic behavior presets with broader built-in risk coverage and CI checks that enforce atomicity and parity with the example behavior specs (#293).
+- Clean-install checks for every documented ASSERT/example environment, including `pip check` and credential-free target imports (#336).
+- Real Docker containment tests that run on every pull request and push to `main` (#328).
 
 ### Changed
 
-- **Breaking, planned for 0.3.0:** optional dependencies now describe ASSERT capabilities rather than repository examples. `otel` is replaced by `phoenix`, `azure-aad` by `azure-auth`, and `embeddings` remains the offline/local embedding backend. The `langgraph`, `dspy`, `examples`, and `regression` extras are removed; install each agent framework from the target project's dependency manifest or the repository example's adjacent `requirements.txt`. `all` now means all optional ASSERT product capabilities: `phoenix`, `analysis`, `embeddings`, `azure-auth`, and `acs`.
+- **Breaking:** optional dependencies now describe ASSERT capabilities rather than repository examples. `otel` is replaced by `phoenix`, `azure-aad` by `azure-auth`, and `embeddings` remains the offline/local embedding backend. The `langgraph`, `dspy`, `examples`, and `regression` extras are removed; install each agent framework from the target project's dependency manifest or the repository example's adjacent `requirements.txt`. `all` now means all optional ASSERT product capabilities: `phoenix`, `analysis`, `embeddings`, `azure-auth`, and `acs` (#336).
 - OpenInference instrumentors are owned by the runtime they instrument. ASSERT installs its LiteLLM instrumentor directly for Prompt Agent tracing; LangChain, OpenAI, and other framework instrumentors live with their target/example dependencies.
 - Analysis keeps NumPy and the OpenAI embedding backend, while Phoenix owns the Pandas dependency used to convert its DataFrame spans.
 
@@ -27,14 +40,20 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 | `assert-ai[examples]` | Each example's adjacent `requirements.txt` |
 | `assert-ai[regression]` | `uv sync --group dev` from a repository checkout; SciPy is no longer required |
 
-- **Breaking, planned for 0.3.0:** `pipeline.inference.max_turns` now resolves to `6` instead of `10` when omitted. Scenario cases that relied on the implicit budget run four fewer turns, which can lower observed violation rates for multi-turn erosion harms, so results from before and after this change are not directly comparable. Set `max_turns: 10` explicitly to keep the previous behavior. Single-turn `prompt` cases do not use this setting and are unaffected. Because the resolved value feeds the inference resume fingerprint, a run interrupted before the upgrade re-executes rather than resumes.
+- **Breaking:** `pipeline.inference.max_turns` now resolves to `6` instead of `10` when omitted. Scenario cases that relied on the implicit budget run four fewer turns, which can lower observed violation rates for multi-turn erosion harms, so results from before and after this change are not directly comparable. Set `max_turns: 10` explicitly to keep the previous behavior. Single-turn `prompt` cases do not use this setting and are unaffected. Because the resolved value feeds the inference resume fingerprint, a run interrupted before the upgrade re-executes rather than resumes (#337).
 
 ### Fixed
 
-- `assert-ai --version` now reads the installed distribution metadata instead of reporting a hard-coded stale version.
-- Declare `aiohttp`, which is directly used by the HTTP endpoint target, instead of relying on LiteLLM to install it transitively.
-- Keep Bank Manager's GPT and non-GPT Azure routes compatible with ASSERT's OpenAI dependency range, and verify its documented installation with `pip check`.
-- Traced callable targets that capture no spans now record an actionable invalid-trace warning instead of silently reporting valid trace metadata.
+- Sandbox policy evaluation, mock resolution, and action evidence can now run in an authenticated host service instead of trusting the evaluated target to report its own decisions (#329).
+- Sandbox mediation now propagates and validates one case ID across policy selection, mocks, execution, and evidence (#327).
+- Proxy-generated network evidence is stored in a host-only directory outside the evaluated target's writable mount (#326).
+- `assert-ai --version` now reads the installed distribution metadata instead of reporting a hard-coded stale version (#334).
+- Declare `aiohttp`, which is directly used by the HTTP endpoint target, instead of relying on LiteLLM to install it transitively (#336).
+- Update LiteLLM to patched version 1.84.1 and cap it below 2.0 (#330).
+- Update vulnerable website, viewer, and Python dependencies to patched releases (#335, #336).
+- Keep Bank Manager's GPT and non-GPT Azure routes compatible with ASSERT's OpenAI dependency range, and verify its documented installation with `pip check` (#336).
+- Traced callable targets that capture no spans now record an actionable invalid-trace warning instead of silently reporting valid trace metadata (#337).
+- Keep viewer result tables within their containers, wrap or truncate long labels, and position metric dropdowns so overflow cannot clip their options (#342).
 
 ## [0.2.0] - 2026-08-14
 
@@ -82,6 +101,7 @@ front-doors.
 - Docs site: rewrite repo-relative `.md` links so cross-references resolve on `responsibleai.github.io` (#225); repair broken links and remove missing logo asset (#227); remove maintainer-agent docs from the public site (#251).
 - Docs / packaging polish: absolute README links for PyPI (#214), static MIT badge (#213), redeploy site when `docs/` or `assets/` change (#212), local viewer guide typos (#208, #210).
 
-[Unreleased]: https://github.com/responsibleai/ASSERT/compare/v0.2.0...HEAD
+[Unreleased]: https://github.com/responsibleai/ASSERT/compare/v0.3.0...HEAD
+[0.3.0]: https://github.com/responsibleai/ASSERT/compare/v0.2.0...v0.3.0
 [0.2.0]: https://github.com/responsibleai/ASSERT/compare/v0.1.0...v0.2.0
 [0.1.0]: https://github.com/responsibleai/ASSERT/releases/tag/v0.1.0

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "assert-ai"
-version = "0.2.0"
+version = "0.3.0"
 description = "YAML-driven safety evaluation pipeline with LiteLLM-backed stages"
 readme = "README.md"
 requires-python = ">=3.11"