The watch floor
Four instruments. One discipline.
Built dark, quiet until a signal is real — one watches the ocean, one watches what comes ashore, one reads a single file, one shows where the answers come from. Pick a door.
SENTINEL
An always-on sector radar and a streaming security feed — thousands of contacts, and the one that earns your attention.
Open SENTINEL → US Import Shipment WatchMANIFEST
The nation’s inbound flow, down to a single port desk — bills, entries, releases, and the one that doesn’t add up.
Open MANIFEST → Knowledge BaseKB demo
KB demo.
Open KB demo → Traveller RecordDOSSIER
One person, one screen — who the record says they are, and how much of it is actually known.
Open DOSSIER → Model EvaluationOTHELLO
A local model taught to stop losing — 66 games against a 1992 cartridge, and the one change that fixed it was not teaching it anything.
Open OTHELLO → Model Evaluation · Part ThreeOTHELLO III
Half the run had measured a model that wasn’t allowed to reason — and what it took to let it think out loud on one machine.
Open PART THREE → Serving Stack27B THINK
The engine behind the games: making a 27B model fast enough to reason on one Mac, from about 35 to 60 tokens a second.
Open THE ENGINE → Vision Model Fine-tuneLAYA vs SF2
A 256M vision model learns Chun-Li from two screenshots and a line of RAM — and the harness that had to be right before any number meant anything.
Open LAYA → Two-System LoopLAYA VS SF2 (Part 2)
A 27B model coaches the small one between rounds — and plain counting beats it by 28 hit points a round. Why, and what the paired tests showed.
Open LAYA Part 2 → How It WorksLAYA VS SF2 (The Loop)
One frame through the machine: the eye reads the screen, three small models vote, memory decides — and between rounds Qwen coaches. Running live.
Open THE LOOP →What this is and why — read the pitch →