Skip to content

CHANGELOG.md

Latest commit

 

History

History
124 lines (102 loc) · 9.48 KB

CHANGELOG.md

File metadata and controls

124 lines (102 loc) · 9.48 KB
Raw

Changelog

All notable changes to CathodeScreen will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

1.4.0 - 2026-03-16

Added

  • Production active learning loop: Closed-loop system connecting MAYBE predictions to DFT/experimental validation, ground-truth feedback ingestion, and automatic retrain triggering. New module active_learning/production_loop.py with FeedbackIngester, ProductionALOrchestrator, and FeedbackPool. Endpoints: POST /feedback, GET /active-learning/status.
  • Multi-chemistry support: Screening extended beyond Li-ion to Na-ion, solid-state electrolytes, and Li-S cathodes. Each chemistry has dedicated composition guardrails, voltage lookup tables, and stability thresholds. Auto-detection from structure composition. Endpoints: GET /chemistry/supported, POST /chemistry/detect.
  • Composition-only fast triage: Sub-millisecond formula screening (no CIF or ML model required) using empirical rules, stoichiometry validation, and voltage/capacity estimates. Supports batch triage of millions of compositions. Endpoint: POST /triage.
  • LIMS integration: Bidirectional integration with laboratory information management systems. Adapters for LabWare, Benchling, and generic webhook LIMS. Inbound webhooks feed DFT/experimental results into the active learning loop. Outbound pushes screening results to LIMS for sample tracking. Endpoints: POST /lims/webhook, GET /lims/status.
  • OPTIMADE API compatibility: Implements OPTIMADE v1.1 specification for interoperability with Materials Project, AFLOW, NOMAD, and other materials databases. Filter parser supports HAS, comparison operators, and CathodeScreen custom properties (_cathode_decision, _cathode_ehull_pred). Endpoints: GET /optimade/v1/info, GET /optimade/v1/structures, GET /optimade/v1/info/structures.

Changed

  • pyproject.toml version bumped to 1.4.0.
  • .ci/empty-file-allowlist.txt updated for new package init files.

1.3.0 - 2026-03-16

Added

  • RBAC (Role-Based Access Control): Three-tier role system (viewer, operator, admin) with fine-grained permissions replacing flat API keys. Backward compatible — existing keys default to operator role. Configured via CATHODE_RBAC_ENABLED, CATHODE_RBAC_KEYS_FILE.
  • Multi-tenancy: Organization-level data isolation via X-Tenant-ID header and per-key org_id binding. Admin keys can act across tenants (super-admin). Configured via CATHODE_MULTI_TENANT.
  • SSO/SAML/OIDC integration: Enterprise single sign-on support with SAML 2.0 and OpenID Connect. JWT session tokens with configurable role mapping from IdP groups. Routes: /auth/sso/login, /auth/sso/callback/*, /auth/sso/metadata.
  • Auth info endpoint: GET /auth/info returns caller identity, role, and tenant context.
  • Kubernetes deployment: Full Kustomize-based manifests under deploy/k8s/ with base, staging, and production overlays. Includes Deployments, Services, HPA auto-scaling, Ingress, NetworkPolicies, PodDisruptionBudgets, and PVCs.
  • HPA auto-scaling: Backend (2-10 pods), Celery workers (1-8 pods), and frontend (2-6 pods) with CPU/memory-based scaling and stabilization windows.
  • Python SDK enhancements: AsyncCathodeClient for async/await usage, predict_and_wait() for blocking async prediction polling, registry and audit access methods, multi-tenant org_id support.
  • ISO 9001 quality management documentation: Full QMS document (CS-QMS-001) covering quality policy, risk assessment, control plans, and CAPA process.
  • IATF 16949 automotive supplement: Automotive-specific compliance document with MSA analysis, control plans, PPAP evidence mapping, and 8D problem-solving alignment.

Changed

  • web/api/main.py: get_api_key now returns Identity objects when RBAC is enabled; added require_permission() dependency factory.
  • sdk/cathode_screen/client.py: Added api_version and org_id parameters, automatic retry transport, enterprise methods.
  • sdk/pyproject.toml: Version bumped to 1.3.0.
  • pyproject.toml: Version bumped to 1.3.0.

Dependencies

  • Added optional [sso] dependency group: pyjwt>=2.8 (optional, fallback to HMAC)

1.2.0 - 2026-03-16

Added

  • Async prediction queue: Celery + Redis based task queue for offloading batch predictions to GPU workers. New endpoints: POST /predict/async, GET /predict/async/{job_id}.
  • GPU dynamic batcher: DynamicBatcher accumulates structures and dispatches optimally-sized batches to minimize GPU idle time. Configurable via CATHODE_ENABLE_BATCHING, CATHODE_BATCH_SIZE, CATHODE_BATCH_TIMEOUT_MS.
  • Shadow deployment: Run a candidate model alongside production on N% of traffic. Compares decisions and logs disagreements for safe promotion. Endpoints: GET /shadow/stats, GET /shadow/analysis.
  • Drift alerting: Automated PSI-based drift detection with multi-channel alerting (Slack webhooks, PagerDuty Events API v2, generic webhooks). Runs hourly via Celery beat.
  • Model registry: Dual-backend registry (local JSON + MLflow) for versioning, governance gating, and stage promotion (staging → production → archived). Endpoints: GET /registry/models, GET /registry/production.
  • Locust load testing: Full load test suite at tests/load/locustfile.py with realistic traffic patterns (single predictions, batches, monitoring, health). Target: 1000 predictions/minute.
  • Celery worker and beat: Docker services for async inference workers and periodic drift monitoring.
  • Redis service: Added to docker-compose for task queue and result backend.

Changed

  • docker-compose.yml: Added redis, celery-worker, celery-beat services with health checks.
  • pyproject.toml version bumped to 1.2.0.
  • Production predict endpoint now fires shadow predictions asynchronously when enabled.

Dependencies

  • Added optional [queue] dependency group: celery[redis]>=5.3, redis>=5.0
  • Added optional [registry] dependency group: mlflow>=2.10
  • Added optional [loadtest] dependency group: locust>=2.20

1.1.0 - 2026-03-16

Added

  • API Versioning: All endpoints now available under /v1/ prefix. Unversioned routes remain for backward compatibility.
  • PostgreSQL audit trail: New CATHODE_AUDIT_BACKEND=postgres option with full schema, connection pooling, and indexed queries. Falls back to JSONL if unavailable.
  • DVC pipeline: Data version control with dvc.yaml defining reproducible stages from fetch → train → calibrate → evaluate → release.
  • Model Card: Formal MODEL_CARD.md following Mitchell et al. (2019) framework with full training data, metrics, limitations, and ethical considerations.
  • Data version pinning: data/DATA_VERSION.json tracks exact Materials Project query parameters, dataset statistics, and data hashes.
  • Integration test suite: End-to-end tests covering prediction pipeline (CIF upload → decision), input validation, API contracts, and audit trail verification.
  • CIF test fixtures: Known-answer materials (LiCoO2, LiMn2O4, LiFePO4) and invalid compositions (NaCl) for regression testing.
  • V&V documentation: Installation, Operational, and Performance Qualification protocols under docs/validation/.
  • Traceability matrix: Requirements → Tests → Evidence mapping for regulatory compliance.
  • CHANGELOG: This file, tracking all versioned changes.

Changed

  • pyproject.toml version bumped to 1.1.0
  • pytest now discovers tests from both src/cathode_screening/tests and tests/ directories
  • Audit trail in prediction endpoints now uses configurable backend selector (get_audit_backend())

Dependencies

  • Added optional [postgres] dependency group: psycopg2-binary>=2.9
  • Added optional [dvc] dependency group: dvc>=3.0, dvc-gs, dvc-s3
  • Added httpx>=0.25 to [dev] dependencies for integration testing

1.0.0 - 2025-01-15

Added

  • Initial release of CathodeScreen
  • 5-member MACE-MP-0 fine-tuned ensemble with quantile regression
  • Conformal calibration for 90% prediction interval coverage
  • Out-of-distribution detection (3-gate: composition, embedding, disagreement)
  • Decision policy with KEEP/MAYBE/KILL classification
  • FastAPI backend with authentication, rate limiting, and CORS
  • Next.js 14 frontend with prediction UI, database viewer, and discovery dashboard
  • Discovery campaign engine with active learning loop framework
  • JSONL audit trail with daily rotation
  • Prometheus and OpenTelemetry observability
  • Artifact manifest with HMAC signing
  • Docker deployment (Render backend + Vercel frontend)
  • GCP Cloud Run deployment support
  • 17 unit test files covering core inference, policy, calibration, OOD, and security
  • SOAP-LOCO validation methodology
  • Governance checks (6/6 automated gates)
  • Cathode property calculators (capacity, voltage, energy density)

Governance Results

  • Test MAE: 0.030 eV/atom
  • Spearman ρ: 0.663
  • Calibration coverage: 91.3% (target 90%)
  • KEEP precision: 92.7%
  • False-kill rate: 0.0%