dakheel/hudanet-v43-code-review

HUDA-Net V43 — Code Review Package (Latest Work)

This repository snapshot is prepared for code review only. It is not a production deployment package and intentionally excludes final holdout datasets, model weights, credentials, and private Hugging Face tokens.

What is included

base_v43/

Frozen V43 application/core source currently used as the reference implementation, including:

• app.py
• hudanet_core/
• build/guard scripts
• requirements
• V43 active contract and deployment/training notes

latest_reranker_work/

Chronological reranker investigation and training work:

1. verified-only training
2. learning-capacity sanity check
3. strong verified training that exposed catastrophic forgetting
4. epoch-by-epoch collapse audit
5. epoch-3 candidate finalization
6. anchor-preserving training
7. anchor weight 0.5 variant
8. anchor weight 0.5 + slower learning rate variant
9. latest approach: full Top-20 ranking preservation while repairing verified errors

Active frozen V43 ranking path

Query -> retriever Top-20 -> reranker -> RRF (k=60) -> confidence/calibration -> visible grounded evidence.

The current investigation is focused on the reranking stage. Candidate-depth experiments showed that simply increasing the number of retrieved candidates was expensive and recovered very few final Rank-1 errors.

Important review boundary

Do not infer final accuracy from the included development experiments. Final/retired holdouts are intentionally not included. The latest full_top20_preservation experiment is a candidate training method, not an accepted production change.

See docs/CODE_REVIEW_NOTES.md for the exact findings and review questions.