cbib/TrialMatchAI: TrialMatchAI 0.9.0 — CLI and reliability foundation

Abstract

TrialMatchAI 0.9.0 establishes a tested CLI and package-delivery foundation, with integrity and recovery fixes from the codebase and Copilot reviews. Try it Use Python 3.11 in an activated virtual environment: python -m pip install trialmatchai==0.9.0 trialmatchai demo --workdir ./demo-workspace trialmatchai demo --workdir ./demo-workspace --resume Open the printed HTML report path. The demo uses synthetic FHIR/trial data, deterministic CPU embeddings and real LanceDB, with model downloads disabled. What changed Added demo, --version, and portable SHA-256 manifest/verification commands, plus strict bootstrap checksums. Changed bootstrap archives now replace managed trees from fresh extraction. Previous trees are retained, unrelated models are preserved, concurrent writers are rejected, interrupted publication can recover, and corrupt caches receive a bounded fresh-download retry. Explicit unresolved FHIR subjects can no longer be assigned to the sole patient, including identifier/display-only, empty, and null references. Demo resume repairs missing/truncated patient reports without rerunning completed ranking. Query-expansion model access is checked before expansion; disabled model stages avoid unnecessary access checks. CI and release publishing share frozen verification, security checks, CPU model imports, package checksums/build inputs, provenance, and an isolated installed-wheel CLI e2e. Documentation builds use pinned Actions and frozen dependencies. Rewritten README and docs landing page with an SVG overview, pipeline diagram, accurate setup instructions, and a sequenced production roadmap. Validation and upgrade notes 474 local tests passed, including real CLI e2e and recovery/attribution regressions. The installed 0.9.0 wheel passed the synthetic workflow, resume/report repair, packaged-catalog, and checksum checks. The base/development advisory audit reported no known vulnerabilities without exceptions. Hosted results are linked from PR #31 and the Release workflow. Strict bootstrap re-extracts legacy markers. Stop readers while updating: publication uses two directory renames and stages publish individually. Keep disk space for retained backups and quarantined downloads. Legacy model roots without ownership metadata are preserved conservatively. Strict resume checks provenance and managed roots, not every extracted file. This is a beta software foundation release. GPU inference and the optional model dependency stack remain unqualified; unresolved advisories and adapter-path compatibility issues are documented. Clinical eligibility semantics, evaluator corrections, complete cache identity, bounded agents, the clinical review workspace, and production application deployment remain roadmap work. Ranking benchmark scores were not rerun or claimed to improve. See the README, release runbook, validation record, and production roadmap. Full changes: https://github.com/cbib/TrialMatchAI/compare/v0.8.2...v0.9.0