← Marilyn

Rabbit holes

Dead ends, with reasons

• A frontier open model as Marilyn’s teacher. Its weights were archived locally as a hedge, never cleaned up. Running it at home managed about a quarter of a token per second. Dead end as a teacher: corpus-scale generation needs an API or rented multi-node serving. Kept as an exact-reference path.¹
• Plan A, learning from context alone. Could she absorb study material just by reading it in the window? Dead end: the Knowing–Using Gap in miniature; nothing routed to answers.²
• A reinforcement-learning phase for the v3 generation. The harness gate for it ran and confirmed the null hypothesis: capability that will not fit in the weights should be reached, not forced. Recorded as a decision record, no RL phase for v3.³
• A phase-D ranking change in the retrieval router. Built on a tag-hygiene result that did not survive a blind set. Withdrawn. ⁴
• A byte-level Marilyn with learned tokenisation. Sketched as a sibling arm to native memory; parked because the vocabulary decision (16k, tied) was settled by measurement first.⁵

Detours that became the next thing

• The nursery. A single serve-and-learn process for a 45M-parameter student with a look-up tool over a small library. Became the seat for every version since, the checkpoint ring, the promotion gate’s canary, and the place where teaching happens.⁶
• “Build her a body.” A sandboxed workflow framework so a model (or its stand-in) can run experiments: read repos, write code, train small models. Decided in July, parked behind hardware, and its containment requirement became the kernel sandbox proven in August. ⁷
• Native memory (Mantis). A weekend sketch of memory inside the model, following the Metis paper. Became the third v5 variant and the self-care programme. See Reading that shaped it.
• Memory tenancy and the “Susan leak”. Reading a memory-pack architecture (Dendritron) for how users’ data could stay isolated when a shared model learns from experience. Became a design with exactly two channels to close: a privacy certificate at the promotion gate and owner-scoped addressing at runtime. A design so far; the scrub gate before consolidation is the first piece to land. ⁸
• An action-state tracker, from a neuroscience model of place cells. Track “where am I in the task” from tool-call actions only, content-free, so the same factorisation gives a privacy boundary, context compression and loop detection. A sketch; its loop-detection half is the one most likely to ship. ⁹
• Concept ladders. A 900-concept dependency graph for a curriculum, 162 rungs built before the corpus work took priority. Became the curriculum gate armed in August.¹⁰
• Retrieval sidesteps the gap. So take the sidestep seriously. The most consequential detour: if knowledge outside the weights is more usable than knowledge inside, build that. Became LiLM, and the reason all three v5 variants are parked awaiting its results.
• The Knowing–Using Gap itself. A paper read on 2026-07-24 that named the failure v2 had just shown. Became the charter’s spine, a weekly review programme, and a consequences plan that decides what changes for each paper read. ¹¹