Lab

Experiments with a written record. Each one asks a concrete question, runs long enough to measure, and publishes the evidence: protocols, reviews, logs, and the scripts behind every number. The record lives at github.com/stierms/lab.

August 2026

Running a local 27B coding agent for 811 minutes: where the time actually goes

A local Qwen3.8-27B on one RTX 4090 built a Rust/Svelte/Tauri Tetris at 35 tokens a second; GPT-5.6 reviewed it read-only; I measured seven rounds. Good code, completion reports that could not be trusted, and a third of the clock spent on compaction.