Journal
Working papers, design notes, and small-pieces-of-reasoning from building a sovereign local-AI stack. Some are formal; some are field-notes. All are dated and revised in public.
We're not in the habit of saving "papers" for "later." We write to think. The journal is here because we want the math and the discipline to be auditable, not because we expect anyone to read all of it.
- Killing a Q5_K Bitplane Bug A full forensic write-up: how the Crimson Analyst toolkit isolated a real, months-hidden Q5_K quantisation bug in a local inference engine — bit-plane comparison, the dequant mismatch, and the one-line fix. The receipt behind the Analyst product page. read →
- Hybrid Retrieval Pipeline The four-stage retrieval architecture behind our code-intelligence suite: BM25 lexical ranking, dense embedding recall, Personalised PageRank over the symbol graph, and Reciprocal Rank Fusion to merge them. Hand-rolled Rust, zero Python in the inference path.
- Deep Research Framework A multi-agent research architecture built around three load-bearing constraints: provenance-anchored claims, opposition evidence treated as equipment rather than stage-prop, and synthesis that never asserts what it can't defend. Five parallel research paths, multi-dimensional viability scoring, and weighted contra-feedback.
- Weaving the Corpus How we assemble training corpora under strict-permissive licensing discipline — per-file LICENSE audits on The Stack v2 (not per-repo, because license-laundering is real) and a provenance hash on every slice, so the lineage of every byte in a model is auditable.
- A Local-First Ternary Model Design notes on our first locally-trained ternary-weight model: Lyapunov-stabilised Liquid Time-Constant state-space layers, BitNet-class compression, and hand-rolled Rust inference small enough to run at conversational latency on edge hardware.
More in flight — papers on Liquid Time-Constant networks, ternary inference kernels, and the engineering of the code-intelligence suite. They land here as they reach "publishable" status — which doesn't mean "polished"; it means the math is real and the language is honest.
Each paper carries a draft-revision number. Drafts can be wrong. If you find a math error, please write to us — hello@scarletsystems.co.nz.