Recognizer comparison versus system comparison
Two different questions must not be mixed.
1. Which recognizer reads a line more accurately?
Answer this only with the protocol-matched CER/WER tables in README.md and BENCHMARKS.md. A model trained on another split cannot be declared worse from a published number measured under another protocol.
2. Which workflow provides more research support?
muharaf_rec_best.mlmodel and exp9 are recognizer weight files. Athar is an application architecture around exp9.
| Capability | Standalone recognizer model | Athar system with exp9 |
|---|---|---|
| One visual transcription | Yes | Yes |
| Preserve raw visual reading | Depends on caller | Yes |
| Up to eight visual alternatives | Not exposed by a plain model call | Yes |
| Local domain-aware LM reranking | No | Yes |
| Optional external-LLM suggestion | No | Yes, advisory only |
| Source-library retrieval | No | Yes |
| Unique versus ambiguous attribution | No | Yes |
| Review queue with reasons | No | Yes |
| Auditable human accept/edit/reject | No | Yes |
| PAGE-XML and TEI export | No | Yes |
| Human-only training-package default | No | Yes |
| Cache/model/source provenance hashes | No | Yes |
This table proves a capability difference, not a numerical accuracy advantage. Numerical recognition claims must come from the benchmark tables.