add honest limitation note (not exact on long chains; Neural GPU drift)

README.md

@@ -33,3 +33,7 @@ covers tiers 1-7. See `manifest.json` for the full architecture and training sum
 - `entry_class`: `model.BitSerialReducer`
 - `output_base`: 2
 - Files: `model.py`, `manifest.json`, `weights.pt`.
+
+## Limitation (honest)
+
+This model passes the random-operand benchmark but is not exact. On structured inputs (powers of two and other long doubling chains) the per-step reduction drifts for some primes beyond about 500 steps, reproducing the Neural GPU limitation (Price, Zaremba, Sutskever 2016). The benchmark tiers reflect average-case accuracy on the official scorer's random-operand distribution, not worst-case exactness of the underlying modular-multiplication operator.