Highlights
- Evidence-first two-stage recommendation system with strict temporal evaluation.
- TRACE four-route multi-interest retrieval with time decay and smooth-max aggregation.
- Constraint-aware diversity reranking, multi-task ranking, anchored calibration, and CPU FastAPI workbench.
- Independent evidence lanes for Amazon ESCI search ranking, Criteo FairJob ads auditing, Hillstrom RCT/uplift, and offline A/B engineering.
- Responsive zero-install portfolio page plus the full local five-view serving workbench.
Measured evidence
- KuaiRec Dev: TRACE NDCG@20 +15.14% and Recall@50 +41.50% versus single-interest; Coverage@50 decreases 10.71%.
- KuaiRand-1K External Final: TRACE NDCG@10 0.39118 versus 0.38531; delta +0.00587 with 95% CI [-0.00983, 0.02312], so the improvement is not statistically conclusive.
- Amazon ESCI: Pairwise LTR NDCG@10 0.80175 versus BM25 0.78427 on 8,850 fixed Test queries.
- FairJob: DeepFM ROC-AUC 0.69863; fairness penalty did not generalize, so lambda=0 remained selected.
- A/B dry run: 20,000 deterministic assignments, SRM p=0.6714, A/A CI includes zero, and the known +4pp effect is recovered as +4.372pp.
Release quality
- Sanitized single-root public history with GitHub noreply identity.
- 365 unit and contract tests pass on Python 3.10 and 3.12.
- Ubuntu and Windows public-snapshot and fresh-wheel matrices pass.
- Immutable External Final report and frozen-config hashes verify.
- GitHub Pages deployment is live.
- Wheel SHA-256: b09390c54ac784b4ba6d8e2f7026a0d1e645f6ad4c13b5cc33803608685f3270
Public boundary
This source/portfolio release excludes raw and processed datasets, checkpoints, model weights, per-user predictions, embeddings, indices, deployment bundles, logs, and private Git history. The historical v2.2.0 Docker verification remains documented evidence; v2.3.0 does not claim a new Docker E2E or a rerun of any Final evaluation.