HUDA-Net v43-AR
Final frozen Arabic neural retrieval release for the HUDA-Net research project.
Frozen architecture
- Stage-3 TOP2 Arabic Retriever
- Stage-4 v2 Arabic Pairwise Cross-Encoder Reranker
- Fixed Reciprocal Rank Fusion,
k = 60 - Stage-5 ranking-confidence calibrator
- Fixed abstention threshold,
0.702353569014631 - Visible evidence grounding only
- Hidden answer pools are forbidden
No model weights, fusion constant, calibration feature order, or abstention threshold were changed after the final blind evaluation.
Final one-shot frozen blind evaluation
The frozen blind split contained:
- 222 locked records
- 191 evaluation queries
- 53 leakage families
- 2796 total Arabic candidate passages
| Component | Recall@1 | Recall@5 | Recall@10 | MRR@10 |
|---|---|---|---|---|
| TOP2 Retriever | 44.50% | 68.06% | 74.87% | 54.64% |
| Stage4-v2 Reranker | 48.69% | 70.68% | 76.96% | 57.43% |
| Frozen RRF pipeline | 45.55% | 70.68% | 76.96% | 56.06% |
Abstention on final blind
- Coverage: 30.37%
- Precision when answering: 82.76%
- False-answer rate when answering: 17.24%
- Correct top-1 recall retained: 55.17%
The development target for precision when answering was 85%. The frozen blind result was 82.76%. This result is reported as-is; the threshold was not changed after seeing the blind result.
Why RRF remains the released architecture
The final blind result shows that the reranker alone scored higher than RRF on Recall@1 and MRR@10. However, RRF was selected and frozen before the blind was opened. Switching to the reranker after seeing blind results would be post-blind model selection and would invalidate the scientific contract. Therefore v43-AR releases the precommitted RRF pipeline.
Scientific safeguards
- Arabic-only training content
- Group/family-separated development splits
- No static semantic keyword rules
- No manual synonym dictionary
- No dialect dictionary for semantic decisions
- No synthetic paraphrases in the final Arabic training stages
- Frozen blind consumed once
- No post-blind tuning
- Visible evidence must equal answer-support evidence
- Hidden answer pool forbidden
Repository layout
retriever/
reranker/
runtime/
runtime_v43_ar.py
config/
ABSTENTION_POLICY.json
CALIBRATION_FEATURE_CONTRACT.json
RANKING_CONFIDENCE_CALIBRATOR.joblib
reports/
STAGE6_REPORT.json
STAGE7_FINAL_BLIND_REPORT.json
release/
RELEASE_MANIFEST.json
Intended use
HUDA-Net is a research decision-support system for Arabic Hajj and Umrah guidance. It should not be treated as an independent religious authority. Runtime answers are expected to remain grounded in the selected visible source evidence and to abstain when confidence is below the frozen threshold.