Fix over-stated quality claims: model is first-sentence-coherent then loops, not "fluent/on-topic"

README.md

@@ -63,11 +63,13 @@ above is the honest held-out figure.
 
 ## What it is good at / not
 
-At 5M parameters this model produces fluent, grammatical, on-topic English,
-but it is a small model: factual recall is weak and greedy decoding drifts
-into repetition. Sampled decoding (top-p 0.9, temp 0.7) is noticeably more
-diverse. It is a demonstration of from-scratch small-model training, not a
-useful general assistant.
+At 5M parameters this model produces grammatical first sentences but is far
+from fluent: greedy decoding is coherent for roughly the first sentence and
+then collapses into repetition loops (e.g. "the world's largest city in the
+world is the world's largest city…"), and sampled decoding (top-p 0.9,
+temp 0.7) is more varied but still drifts into incoherence within a few
+sentences. Factual recall is weak. It is a demonstration of from-scratch
+small-model training, not a useful general assistant.
 
 ### Sample outputs (greedy, temp 0.0)
 
@@ -93,7 +95,7 @@ useful general assistant.
 
 ## Version
 
-This is **v2** of CompactLM-5M. It replaces an earlier 6,162,688-param build (d256/4L/4H, vocab 12288) whose card honestly noted it produced grammatical word-salad. This v2 retrain (d224/6L/GQA, vocab 8192) is the first version of this lineage that produces genuinely coherent, on-topic output, so it is published as the canonical model.
+This is **v2** of CompactLM-5M. It replaces an earlier 6,162,688-param build (d256/4L/4H, vocab 12288) whose card honestly noted it produced grammatical word-salad. This v2 retrain (d224/6L/GQA, vocab 8192) produces grammatical first sentences (better than the v1 word-salad) but still collapses into repetition loops under greedy decoding, so it is published as a from-scratch training demonstration, not as a coherent generator.
 
 
 - `model.safetensors` — weights (27 MB, F32). `head.weight` and `tok.weight`
@@ -106,4 +108,4 @@ This is **v2** of CompactLM-5M. It replaces an earlier 6,162,688-param build (d2
 
 Trained with a custom GQA LLaMA script (6 layers, d224, SwiGLU). The exact
 script, tokenizer and training log are not bundled here; the architecture is
-fully specified in `config.json` and above.
+fully specified in `config.json` and above.