Research

Testing Clef Decision Models as a Content Filter

I spent four days running Cloudflare's Clef-Flash decision model on a laptop GPU to filter sensitive questions out of a four-million-row dataset. I compared serving engines and ways of asking, and about 650,000 requests showed how much a decision model's answers depend on wording and on the questions asked alongside them.

Testing Blender Scenes Built by LLM Agents

An agent building a Blender scene can't see the viewport, so every claim it makes about clipping, contact, scale or motion has to be measured. These are the twelve families of checks from a 640-frame mech film a team of Claude agents built in bpy, plus the three times the checks themselves were wrong.

How Claude Code's Projects Beta Works Under the Hood

A project is a channel, a thread is a message, and the Claude doing work on your machine is a child of a cloud session you never see. A full walkthrough of the Projects beta, from the channels API down to the process the desktop app spawns in your folder.

DIY Public-Records Research

How to fact-check a story with free Canadian public records: lobbying registries, corporate filings, contracts, building permits, and the access-to-information trail. No paywall, and no press pass needed.

Claude Desktop Cowork Mode: VM Architecture Analysis

The complete technical reference behind the Cowork deep-dive: IPC APIs, refactored symbol mappings, telemetry events, VM image internals, and the full layered security architecture.

Reverse Engineering Claude Desktop's Cowork Mode: A Deep Dive into VM Isolation and Linux Possibilities

Claude Desktop's Cowork mode only runs on macOS. I pointed Claude at the minified app to reverse engineer why, and found a lightweight Linux VM, layered sandboxing, and a real path to Linux support.