LiLM / Lucky Labs

status

active · ongoing

when

2026-08 → now

came from

Marilyn, Fleet & operations, Agent Zone

A research question with a product face: can a model learn, self-learn and be patched without full retraining, if its knowledge lives outside the weights?

The question

Every way Marilyn has of learning something new ends in the same place: a training run. Pre-training, fine-tuning, the nursery’s consolidation, even native memory once it is trained, all change weights, all cost GPU time, and none can be audited afterwards to say what was learned or from where. Correcting one fact means another run. And what is in the weights is not reliably usable even when it is there, which is the Knowing–Using Gap the whole lab is built around.

LiLM asks whether there is another way. Separate knowledge from reasoning. Keep knowledge as explicit, human-readable, content-addressed nodes in a store on fast NVMe, where a fact can be added, corrected, retired or traced to its source without touching a weight. Let a small local model do the reasoning over a sparse retrieved context. If that works, “training” a model on new material becomes ingestion, “self-learning” becomes the model minting and gating its own nodes from what it reads, and “patching” becomes editing a node, with the provenance kept. If it does not work, that is worth knowing before the next 40B-token run.¹

Why now

Because the three v5 variants are the alternative, and they are parked until this answers. The knowing directory experiments inside Marilyn had already shown a node could be minted from real text, gated, repaired and retired; LiLM is that idea built as a system rather than a lane.

What it is

A pipeline of harvest → ingest → verify → query, drawn on Architecture. Documents become semantic nodes with stable content-addressed ids, domain classification, and a full-text index. An ingestion factory built in one autonomous week runs it as Temporal workflows with fail-closed gates, a deterministic build id, atomic ship-and-rollback, and a proven kill-and-resume. Today it indexes the same corpus that trains Marilyn, so the comparison is fair: the same knowledge, inside the weights or outside them. It has a product face, Lucky Labs Co at luckylabsco.io, because a knowledge store that can be patched without retraining is useful to people who will never train a model. Lucky Labs is a hobby brand for that face, not a company with employees or customers.²

Where it stands

Active. The factory merged on 2026-09-02 after a real-corpus gate (the numbers are on Wins). The same gate showed the summariser truncating rather than summarising, and relations between nodes about 100× sparser than on synthetic data (Losses). Decision: ship a searchable substrate as v1; real summarisation and a relations linker are v2.³

Lineage

• Came from: Marilyn (the Knowing–Using Gap, and every learning path ending in a training run), Agent Zone (knowledge written where a machine can read it), fleet ops (everything runs as workflows on the fleet)
• Led to: open. Its answer decides which, if any, of Marilyn’s three v5 variants trains next.

Sub-pages

• Architecture: the factory from document to served build, the query path, and the layered tenant stack
• Stages, measured: every stage from the synthetic million to the real corpus, with the number each one produced
• Wins · Losses · Rabbit holes