Berna R8: Bounded Forgetting in Cell-Based Language Models (Proof of Concept)

Berna R8 is a proof-of-concept follow-up to Berna R7 v3. It fixes five architectural issues identified in R7 v3 (per-cell K5 freshness, per-cell K6 encompassment, per-cell K2 breadth, Plexus threshold, and full DNA persistence) and implements Bounded Forgetting — the central claim of the parent theory R5 (DOI: 10.5281/zenodo.23015308). The H2 experiment demonstrates for the first time that fine-tuning Berna on a new language (German, 2000 steps) improves the target language (DE loss: 4.7406 → 4.0797) without degrading the original language (ENG loss: 4.0781 → 4.0703). The forgetting metric F_eng = max(0, ENG_after - ENG_before) = 0.000000, satisfying the R5 bounded-forgetting criterion (F_j ≤ 1e-6). The R8 model weights and code are available at https://huggingface.co/BernaLabs/berna-r8.