← LiLM / Lucky Labs

Rabbit holes

The experiments inside Marilyn that became LiLM

• The knowing directory. A registry of “knowings”, each a node with provenance, tags, a reachability canary and a maintenance loop: detect, triage, repair or dispose, with no stage skipped and a tombstone that blocks re-minting the same content. Verified by tamper: both gates failed on a one-token change, as they should. Became LiLM’s node model and its fail-closed gates. ¹
• Minting a knowing from real text. The first node minted from a real document bound at +0.97 nats (bootstrap interval 0.36 to 1.57 over the probe set) and earned its existence on four of five bars, with the fifth an honest fail. Binding was five to nine times weaker than hand-written nodes, reported and not interpreted. Became the ingest-then-gate shape of the factory.²
• Lexical routing over tags. A read-only, calibrated router that mounts nodes by tag match. Its measured gaps are the v2 backlog: no stemming, so scrub never matches scrubs; a tag cap of 24 that does hygiene work by accident; and disclosure of what was mounted, which works at small scale and is unlicensed at large. Dead end as a general router; alive as the cheapest thing that works. ³
• Inflated multiples, corrected. Early router numbers (106× and 132× speed-ups) were re-measured with committed instruments at 45× and 49×; the first instrument had never materialised a result set. The conclusions stood; the numbers did not. Recorded as the reason every LiLM figure now cites a committed instrument. ⁴

Detours within LiLM

• A hierarchical hand-off to a second model. The remediation spec for the user stack (twelve findings) was handed to a different vendor’s model while sub-agent waves waited on my signal. Outcome not yet cited on the site: the work is in flight.⁵
• A SQLite-per-cell index. The first substrate. Dead end at full-corpus scale: the per-cell tables were the merge bottleneck. Became flat sorted postings with a Rust merge.⁶
• Build the index on the big GPU. A proof of concept, conditional go, that turns an hours-long merge into seconds of GPU time on hardware that is otherwise idle between training runs. Whether it ships depends on the parked-variants decision.⁷

The one that decides everything

• Can a model learn, self-learn and be patched without full retraining? LiLM exists to answer that, and the answer decides Marilyn’s three parked v5 variants. If an auditable store outside the weights lets new material be ingested, self-minted and corrected without a training run, and the small model can use it, the next tranche of GPU time goes there; if not, one of the three variants gets it. Either way the answer will be a measurement, and it will be on this site.