Wins

The architecture survived real data

The real-corpus gate passed on the substrate. 120,934 documents became 244,242 nodes across five tiers, ingested through the production path in 31.9 seconds. The full-text index answered a cold query in 37 ms at 8,691 queries per second from a 55 MB index, 6% the size of the store it indexes, with 10 of 10 matched-set queries correct and no term lost across 959,067 of 959,068. Coordinate scoping worked. ¹
Deterministic builds, proven byte-identical. A rebuild of the same input produced the same build id, after a real non-determinism (random index UUIDs plus hash-map ordering in two languages) was caught and fixed with a single-threaded, sorted, canonical writer.²
Durability, proven by kill. The build workflows were kill-9’d mid-run and resumed. Shipping is an atomic symlink swap with fail-closed validation and rollback, sabotage-tested and concurrency-safe. The end-to-end loop ran for real: version A shipped, version B built additively on top, rolled back, version A re-served. ³
Four fail-closed gates in the builder, including a term-quality gate, and a QA gate that activates a build only after a discriminative smoke passes.⁴

The whole ingestion factory, phases 0 through 5, was planned, built by sub-agent waves, audited by a PM session and merged in the week of 2026-09-02. I set the plan, the gates and the autonomy limits and reviewed each phase. Five full builds failed before one succeeded, and every one was diagnosed by me: an activity that did not heartbeat on full-size shards, a workspace on the filesystem the kubelet monitors, a parallel pool with no per-shard progress, a storage I/O error on a store with no redundancy, and a database lock timeout sized for a fast disk on a slow one. The agents wrote the code; the failures are on Losses, and reading them is the job. The plan-of-record, the status doc and the end-to-end report are all in the repo. ⁵
The corpus is Marilyn’s. The factory ingests the same composed v5 corpus that trains her, 325 training shards verified byte-identical by checksum, with the validation shards excluded on purpose so retrieval can never contaminate her held-out set.⁶

A flat-postings redesign replaced per-cell SQLite tables with node-grouped sorted runs and a Rust k-way merge, after a five-iteration loop of implement, adversarial review and live 40-shard validation. ⁷
A GPU index proof of concept built the full-corpus index on the training card in about 18 seconds of GPU time, with correctness proven independently: zero ordering violations across 1.68 billion postings. Conditional go.⁸
A weighted-AND shortcut was measured and dropped everywhere once real vocabulary showed it slower than the plain path (0.83× at k=10). ⁹

Lucky Labs Co, at luckylabsco.io: a hobby brand for the part of LiLM that people who never train a model could use. ¹⁰

Numbers on this page are quoted from the lab's own records. The records are private; each note gives the record's date.