TinyRustLM Runtime Source Integration Notes

The uploaded Rust source files make the ModelBreeder theory more concrete. They demonstrate a local-first, browser-oriented runtime that can load a custom .slm model package, parse tokenizers, run quantized matrix-vector operations, apply validated adapter-delta packages, generate text, emit diagnostics, and run local eval cases.

Runtime concepts promoted into the public site

• .slm model packages should expose header identity, tokenizer section identity, tensor directory identity, parameter count, quantization mode, and artifact checksum.
• Adapter packages should not be treated as generic blobs. They should bind to an expected model identity, tensor layout checksum, tokenizer checksum, and parameter count.
• A browser-local runtime can expose strict ABI functions for loading models, validating adapters, applying adapters, configuring sampling, generating text, and reading diagnostics.
• Evaluation should produce sidecars that name the model, cases, quality boundary, evaluator, source kind, and pass/fail result.
• Diagnostics should include model-loaded state, token counts, tokens per second, peak scratch usage, KV cache length, active quantization, applied adapter count, and assembly-state checksum.

Site content updated from this evidence

• /architecture/tinyrustlm-runtime
• /architecture/slm-model-package-abi
• /blueprints/rust-browser-model-lab
• /tools/release-packet-builder
• /tools/lineage-dag-viewer