| # RegalFire scientific product selection | |
| 12 candidates; 5 selected. Scores are editorial assessment informed by actual record quality and 2026 scientific RAG evaluation work, not client commitments. Higher build-cost efficiency means lower implementation cost. | |
| | Candidate | Score /100 | Selected | Actual records | Decision | | |
| |---|---:|---|---:|---| | |
| | Scientific-Citation-Graph | 82 | True | 888 | 888 mechanical reference edges with literal context and full source attribution. | | |
| | Scientific-Answer-Ranking | 80 | True | 73 | 73 genuinely multi-answer threads; community signals explicitly separated from truth. | | |
| | Scientific-RAG-HardCases | 80 | True | 258 | 258 property-selected cases; no invented correct-answer label. | | |
| | Scientific-Code-and-Analysis-QA | 79 | True | 72 | 72 threads and 174 exact preformatted scientific code blocks; no execution/correctness claim. | | |
| | Scientific-Statistics-QA | 71 | True | 89 | 89 source-grounded QA threads; statistical methods/retrieval pilot. | | |
| | Scientific-Computational-Methods-QA | 69 | False | not built | Valid source exists, but a separate site-only QA duplicates the more distinctive code product. | | |
| | Scientific-Methodology-QA | 57 | False | not built | No precise deterministic methodological-topic subset defined; avoid vague ontology labels. | | |
| | Scientific-DataScience-QA | 34 | False | not built | Useful fit but collection not started: shared API quota 12, reserve 10. | | |
| | Scientific-Chemistry-QA | 33 | False | not built | No independently approved question/answer/revision corpus. | | |
| | Scientific-Physics-QA | 33 | False | not built | No independently approved question/answer/revision corpus. | | |
| | Scientific-Astronomy-QA | 31 | False | not built | No independently approved question/answer/revision corpus. | | |
| | Scientific-Math-QA | 30 | False | not built | Large public source, but no verified complete corpus in this run. | | |
| Research signals: [Scientific QA RAG, LREC 2026](https://aclanthology.org/2026.nslp-1.3/), [SciRet 2026](https://arxiv.org/abs/2608.03860), [T2-RAGBench, EACL 2026](https://aclanthology.org/2026.eacl-long.8/). | |
| The family shares a 288-thread corpus. Selected product counts cannot be added as independent source volume. Public holdouts are not contamination-resistant. | |
| Shared API remaining quota was 12 at the last successful native collection. Reserve 10; deferred unverified domains instead of weakening source/license checks. | |