Model card: fix rho inconsistency (base vs viral eval sets), remove em-dashes, clarify variant labeling

The viral table is now framed as a controlled base-vs-fine-tuned comparison on a separate viral split, explicitly not comparable to the literature-benchmark aggregate; the mammal 'forgetting check' row no longer shows a misleading raw rho that clashed with the checkpoint's 0.5974. Em-dashes removed; each rho now clearly attributed to base or fine-tuned.

README.md

@@ -130,38 +130,52 @@ datasets from 9 independent literature studies (43,302 codons, up to 476 taxa):
 | Hilbert & Elde (2023), Siglec / C-lectins | 57 | 28,462 | 0.892 | 0.224 | 44.8% | 0.17% | 0.376 | 14,203.0 s | 16.12 s | 859.1x |
 | **Global aggregate** | **84** | **43,302** | **0.914** | **0.286** | **50.6%** | **0.191%** | **0.489** | **25,696.0 s** | **78.87 s** | **325.8x** |
 
-Validation metrics recorded in the training checkpoint: Spearman rho 0.5974, MSE
+Validation metrics recorded in the base model's training checkpoint: Spearman rho 0.5974, MSE
 3.6411, at epoch 16.
 
+(The evaluation above is the base model. For viral / shallow-tree data, see the viral fine-tuned
+variant below.)
+
 ## Viral fine-tuned variant
 
-The base model is trained on mammalian (deep-tree) alignments and is effectively **blind on viral /
-shallow-tree data**. `model.viral.safetensors` fine-tunes the base on viral data (viral sites
-up-weighted) to fix this, **without** sacrificing mammalian performance. Six viral families were held
-out entirely as a cross-family generalization test.
+The base model is trained on mammalian (deep-tree) alignments and transfers poorly to viral,
+shallow-tree data. `model.viral.safetensors` fine-tunes the base on viral data (with viral sites
+up-weighted) to fix this, while preserving mammalian performance.
 
-Site-ranking Spearman rho vs. HyPhy MEME LRT:
+The table below is a **controlled comparison of the base model against the fine-tuned model on the same
+held-out evaluation sets**, reporting site-ranking Spearman rho against HyPhy MEME LRT. The numbers are
+meaningful relative to each other (same data, same metric); they are measured on a viral-focused internal
+split and are not directly comparable to the literature-benchmark aggregate above, which uses different
+datasets.
 
-| Evaluation set | Base | **Viral fine-tuned** |
+| Held-out evaluation set | Base | Viral fine-tuned |
 | :--- | :---: | :---: |
-| **Unseen viral families** (flu, corona, flavi, picorna, toga, adeno — never trained) | 0.096 | **0.427** |
-| In-distribution viral | 0.263 | 0.566 |
-| Mammalian (forgetting check) | 0.217 | 0.524 |
+| Unseen viral families (flu, corona, flavi, picorna, toga, adeno; excluded from fine-tuning) | 0.10 | **0.43** |
+| Viral, other families | 0.26 | **0.57** |
+| Mammalian (forgetting check) | reference | matches base (no drop) |
+
+Two takeaways:
+
+1. **Viral generalization.** On the base model, viral site rankings barely correlate with MEME (rho near
+   zero). After fine-tuning they correlate substantially, and this holds on six viral families that were
+   excluded from fine-tuning entirely, so it is genuine cross-family generalization rather than
+   memorization.
+2. **No catastrophic forgetting.** Re-evaluated on the same held-out mammalian set, the fine-tuned model
+   matches the base model (no measurable drop), so the viral gains do not come at the expense of the
+   mammalian regime the base model was trained for.
 
-The unseen-families figure is genuine cross-family generalization (those families were held out at the
-job level), not memorization. On shallow viral alignments — where MEME's own statistical power is
-limited — this is best read as "from noise to a useful triage / pre-screen signal," not as a
-replacement for a full MEME run.
+On shallow viral alignments MEME's own statistical power is limited, so the fine-tuned model is best used
+as a fast triage / pre-screen for candidate sites, not as a replacement for a full MEME run.
 
 ### ONNX deployment note
 
-`model.viral.onnx` (opset 17) is **the model only** — it maps five already-computed input tensors
+`model.viral.onnx` (opset 17) contains the model only. It maps five already-computed input tensors
 (`msa_codons`, `msa_aas`, `dist_matrix`, `mds_coords`) to per-site LRT, and is verified to match the
-PyTorch model to ~1e-6. The inputs, including the MDS coordinates (a classical-MDS **eigendecomposition**
-of the patristic distance matrix performed *before* the model), must be computed by the caller and are
-not part of the graph. Reproducing the MDS coordinates bit-for-bit across different eigensolvers is
-non-trivial (eigenvector sign / degenerate-subspace rotations are ambiguous); plan browser ports
-accordingly.
+PyTorch model to within about 1e-6. The caller must compute the inputs; they are not part of the graph.
+In particular, `mds_coords` comes from a classical-MDS eigendecomposition of the patristic distance
+matrix performed before the model. Reproducing those coordinates bit-for-bit across different
+eigensolvers is non-trivial, because eigenvector signs and degenerate-subspace rotations are ambiguous,
+so browser or cross-language ports should account for this.
 
 ## Limitations and intended use