Drop unreproducible design-distance multiple; correct undatabased non-host rate to 99%

README.md

@@ -213,9 +213,9 @@ the hardest identity task at 0.26, stays above; missense pathogenicity, the stro
 threshold, function tasks (red) below, with DNA and protein on the same axis.*
 
 Generation makes the separation concrete. Coordinate ascent on the host head designs sequences with
-host_score up to 22, against a natural ceiling of about 6, and these designs sit in a region of
-composition space roughly 3.6 times the natural spread away from any real sequence, a region that
-maximizes the detector but holds no biology. The 8B language model rates those same designs as
+host_score up to 22, against a natural ceiling of about 6, and these designs sit far outside the
+region of composition space occupied by any natural class, where the detector saturates but no
+biology lives. The 8B language model rates those same designs as
 increasingly unnatural, so composition and neural naturalness are separable axes rather than two views of
 one thing. `ADVERSARIAL.md` covers the order-(k-1) evasion boundary and why a language model resists it;
 `TOOL.md` covers footprint, throughput, read length, and operating modes.

TOOL.md

@@ -42,7 +42,7 @@ conservative retention for enrichment).
 
 - **Strong:** human against bacterial and viral sequence, the clinically dominant contrast.
 - **Reference-free advantage:** on sequence absent from every database, Kraken2 classifies 0% and
-  BLAST 6.6%, while this filter calls 100% (97% as non-host). It is the only option when the
+  BLAST 6.6%, while this filter calls 100% (99% as non-host). It is the only option when the
   sequence has no database match, for example environmental or divergent material.
 - **Not for:** discriminating closely related mammals (human vs mouse/rat is weak by composition);
   use it for host-vs-microbe, not for separating vertebrate species.