Release TrialMatchAI 0.9.1

CHANGELOG.md

@@ -6,7 +6,22 @@ All notable changes to TrialMatchAI are documented here. The format follows
 
 ## [Unreleased]
 
+## [0.9.1] — 2026-09-19
+
+### Added
+- A checksum-pinned, CPU-only `reproduce-paper` command audits the published
+  TREC 2021/2022 archive against official qrels and separates stored metrics,
+  ranking recalculation, retrieval recall, and current-evaluator results.
+- `trec-evaluate` reuses completed rankings to compare explicit unjudged-trial
+  policies, with macro means, medians, per-topic metrics, qrels checksums, and
+  evaluation-input fingerprints.
+- Reproducible reports record both unjudged policies for six complete local
+  configurations across all 125 TREC 2021/2022 topics.
+
 ### Fixed
+- Paper archive caches and fresh extractions must contain every summary,
+  per-topic metric, ranking, candidate list, and recall-cutoff file consumed by
+  the audit.
 - Eligibility assessment is controlled by `rag.enabled` independently of the CoT
   prompt setting. Disabling `use_cot_reasoning` now selects direct JSON assessment.
 - Ranked output records assessment controls and availability. Reports explicitly
@@ -19,6 +34,13 @@ All notable changes to TrialMatchAI are documented here. The format follows
 - The flowchart places eligibility assessment on the default path and labels the
   explicit retrieval-only bypass. Both SVGs state its default-on behavior.
 
+### Changed
+- The TREC package imports its GPU runner lazily so artifact and metric audits do
+  not load the model stack.
+- Paper-reproduction output reports current tie-aware nDCG under both supported
+  unjudged-trial policies while retaining the previous condensed result field for
+  compatibility.
+
 ### Migration
 - Set `rag.enabled: false` to skip assessment. `use_cot_reasoning: false` alone no
   longer disables it. No new GPU or clinical-accuracy qualification is claimed.

README.md

@@ -32,7 +32,7 @@ it does not install the optional model stack.
 ```bash
 python3.11 -m venv .venv
 source .venv/bin/activate
-python -m pip install trialmatchai==0.9.0
+python -m pip install trialmatchai==0.9.1
 
 trialmatchai --version
 trialmatchai demo --workdir ./demo-workspace
@@ -78,8 +78,8 @@ The code calls its generated eligibility explanations **CoT reasoning**; these
 are model outputs to inspect alongside source evidence, not verified clinical
 reasoning or a guarantee that every criterion has been covered.
 
-In source checkouts after 0.9.0, eligibility assessment and CoT prompt style have
-independent switches (the change is currently unreleased):
+Starting with 0.9.1, eligibility assessment and CoT prompt style have independent
+switches:
 
 | `rag.enabled` | `use_cot_reasoning` | Result |
 | :--- | :--- | :--- |
@@ -127,7 +127,7 @@ hardware you select; there is no universal single-GPU capacity promise.
 For the default model pipeline, install its extras in a dedicated environment:
 
 ```bash
-python -m pip install 'trialmatchai[llm,gpu,entity]==0.9.0'
+python -m pip install 'trialmatchai[llm,gpu,entity]==0.9.1'
 ```
 
 The optional inference stack has unresolved dependency advisories and has not

docs/index.md

@@ -18,7 +18,7 @@ criterion-level eligibility assessment. Query expansion is separately configurab
 Install with Python 3.11 in an activated virtual environment:
 
 ```bash
-python -m pip install trialmatchai==0.9.0
+python -m pip install trialmatchai==0.9.1
 trialmatchai demo --workdir ./demo-workspace
 trialmatchai demo --workdir ./demo-workspace --resume
 ```

docs/pipeline.md

@@ -102,7 +102,7 @@ See the [API reference](api.md) for `StageContext`, `Stage`, `select_stages`, an
 
 ## Assessment modes
 
-This section describes the unreleased changes on `main` after 0.9.0.
+This section describes the assessment behavior introduced in 0.9.1.
 
 Eligibility assessment is enabled by default. `rag.enabled` controls whether it
 runs; `use_cot_reasoning` selects the CoT prompt (`true`) or direct JSON prompt

pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "trialmatchai"
-version = "0.9.0"
+version = "0.9.1"
 description = "AI-driven patient-to-clinical-trial matching: hybrid retrieval + LLM eligibility reasoning."
 readme = "README.md"
 requires-python = ">=3.11,<3.12"

src/trialmatchai/__init__.py

@@ -1,3 +1,3 @@
 from __future__ import annotations
 
-__version__ = "0.9.0"
+__version__ = "0.9.1"