Lab record

SeldomSolemn

I build tools for working with local agents and testing ideas. This is the record of what works, what fails, and what comes next.

Projects

Tools and experiments, with notes from the work behind them.

A phone-first remote for local coding agents. Keep working with the agent on your machine when you’re away from the desk.

Disconnect lab interface with a chat workspace and Grok agent connection panel.

Accounts work in lab · computers not connected yet

Screenshot and latest update →

An experiment comparing language models, rule-based strategies, and buy-and-hold on US equities, with trading costs included.

July 25 control run Original reported medians · $100 start
StrategyBalance
Buy & hold SPY$105.93
Mean reversion$81.93
Random$77.37
Legacy momentum$32.01
316 overlapping windows · 180-day horizon · earlier engine with known defects, described in the report Results, replay and lessons →

Updates

Progress notes from the projects I’m working on.

  • Disconnect Build-stage accounts now cover email, sign-in, profile, sessions, and password reset from a mailed link. An account does not open a private agent or connect a computer. Next: install on a phone, then a computer companion. Still in lab; not open for public signup. Read the project →
  • AI Trading Arena Documented the decision loop from the engine: cached daily JSON, the nine-lens snapshot, control screens, the spread-and-stop cost cage, and that models did not see the book or call a quant toolkit. Read the methods →
  • AI Trading Arena Added all six models and four controls to the shared-period performance charts, with hover labels to identify each line. Reworked the research write-up around the question, methods, results, execution costs, drawdowns, and proposed next run. Explore the results →
  • This site The site is live at seldomsolemn.com, a home for project updates, experiment results, and longer writing.
  • Disconnect Polishing connections beyond Grok CLI. Codex CLI, Claude Code, and local models that support ACP are in the pipeline. Still in lab; not publicly shipped.