ILM and ArfaLM: Child-Scale Language Models Trained on Natural Parent-Child Speech for Deployment on Sub-1GHz Legacy Hardware. Part 2: Deployment, Optimization, and Safety on Original Hardware

This is Part 2 of a two-part work on child-scale language models for legacy hardware. Part 1 (Shah, 2026a; DOI: 10. 5281/zenodo. 19605032) reported the training, evaluation, and architectural comparison of ILM (2. 3M parameter LSTM) and ArfaLM (8. 4M parameter Transformer), both trained on CHILDES Eng-UK conversational data. This paper reports three deployment contributions: (1) a KV-cache implementation in ANSI C89 that reduces ArfaLM inference from O (n²) to O (n) per generated token, achieving a measured 26x throughput improvement (from 0. 1 to 2. 6 tokens per second) on the target Pentium II hardware, exceeding the 18x projected from arithmetic analysis; (2) deployment of both models on an original IBM 300PL (Pentium II 400 MHz, 128 MB SDRAM) running Windows 98, compiled with Borland C++ 5. 02 (1997), with measured throughput of 9. 1 tokens per second for ILM and 2. 6 tokens per second for ArfaLM; and (3) an external safety filter comprising a 162-term whole-word blocklist, PII pattern detection, hardcoded greeting and farewell responses, and basic arithmetic capability, with a measured false positive rate of zero on the 5, 000-word CHILDES vocabulary. The safety filter validates the architectural argument from Part 1 that at sub-10M parameter scales, external guardrails applied to curated data and to inputs and outputs at inference time are the more defensible safety strategy than attempting to train alignment behaviour into the model itself. Artifacts: - Part 1 preprint: https: //doi. org/10. 5281/zenodo. 19605032- Dataset: https: //huggingface. co/datasets/nshah-fbcs/childes-engUK-conversational-pairs- Models: https: //huggingface. co/nshah-fbcs/ilm-arfalm-childes ILM is dedicated to the memory of Imran Shah (1969 to 2019). ArfaLM is named for Arfa Karim Randhawa (1995 to 2012). Technology is for everyone.