Lab · Built in the open

Self-directed research,with the receipts

Work I do on my own time, held to the same standard as the paid kind: named baselines, stated uncertainty, and the dead ends published next to the wins.

In active development · Research agent

Beane: small models, better systems

A deep-research agent that turns one question into a long, cited report - every quote located on its page in code before a model sees it. It runs on mid-tier models only, so every gain has to come from how the work is split, checked and merged rather than from the model doing the writing.

LangGraph · Measured on DeepResearch Bench II · Private repo, open source intended

63.0 mean · 2nd on the board

Pre-submission · 1.4 off the lead

Above every frontier-lab product

Read the write-up →

Published study · Currency forecasting

Which currency questions are actually answerable?

Ten years of daily prices for the euro, yen and rupee against the dollar. Three questions, each graded against a stated simple guess. Two were dead ends. The third fixes a number companies set every month.

Forecasting · Public data only · Fully reproducible repo

Q1 · No better than guessing

Q2 · 51-57% less error

Q3 · 16-28% less error

Read the study →

Build log

What landed, and when

2026-09-18

Beane reached a 63.0 mean on DeepResearch Bench II, up 6.7 on the same tasks - recall carried it, up nine and a half points. Second on the board, 1.4 behind the leader and clear of every deep-research product a frontier lab ships. Pre-submission; every label is in the repo ledger.

2026-08-06

Currency study published: three questions, ten years of data, every answer graded against a stated simple guess. Two dead ends documented alongside the one usable result.

2026-07-31

Researcher agent committed. Plan: LangGraph agent, trace recording from day one, replay player, then live mode behind hard caps.

Want to know when something new lands here? Email[email protected]with the subject "lab".