Clarify training and evaluation provenance

Remove an unsupported held-out-prime implication. Training used independently sampled random moduli; separately generated cases served as selection and qualification gates. Model code and tensors are unchanged.

README.md

@@ -55,7 +55,7 @@ local case sets, not a Playground score and not a proof of exactness. The
 Playground run is intended to measure transfer on its test set.
 
 Before publication, the packaged float32 inference path was also run through
-the current official interface on 110 held-out cases. It was deterministic and
+the current official interface on 110 separately generated cases. It was deterministic and
 scored 99/100 across tiers 1-10: 10/10 on tiers 1-9 and 9/10 on tier 10. Tier 0,
 which is an unscored pure-multiplication diagnostic, was 6/10. This was a local
 interface check, not a SAIR Playground result.

manifest.json

@@ -3,5 +3,5 @@
   "output_base": 2,
   "framework": "pytorch",
   "model_description": "A 126,603-parameter, two-layer bidirectional GRU transition cell conditioned on the binary modulus. The cell predicts one binary residue-state update and is reused inside a fixed bit-serial Horner schedule to reduce both operands and multiply their residues. The model emits MSB-first base-2 digits and supports state widths through 2048 bits.",
-  "training_description": "The shared transition cell was trained on synthetic one-step targets s_next = (2*s + d*x) mod p, beginning at L=128 and continued through L=256 and L=512 with fixed seeds and recorded checkpoint gates. The L=512 checkpoint passed an L=1024 update-zero qualification without changing any tensor. Evaluation primes were not used as training targets. This repository is a development research snapshot, not an exactness claim."
+  "training_description": "The shared transition cell was trained on synthetic one-step targets s_next = (2*s + d*x) mod p, beginning at L=128 and continued through L=256 and L=512 with fixed seeds and independently sampled random moduli. Separately generated case sets were used as checkpoint-selection and qualification gates. The L=512 checkpoint passed an L=1024 update-zero qualification without changing any tensor. This repository is a development research snapshot, not an exactness claim."
 }