CALL side is calibrating well. PUT side is underperforming (43% win rate, n=7) — trimming exposure there until the pattern resolves. Numbers update weekly.
I build AI systems for a living. I've traded my own capital for twenty years across US and Indian markets. coworker1 is the workbench where those two practices meet — a problem I know viscerally, being rebuilt with tools that only recently became practical for one person to use end-to-end.
Enough to recognize most of the failure modes when they show up again. Enough to know which tools lie, which data feeds mislead, which calibration loops are missing. Enough to have lost money on trades I'd later recognize as obvious — if the right signals had been in the right order at the right time.
Every trade has a ground truth. Win or loss. Measurable in dollars, in time, in calibration error.
If the 60/30/10 idea is directionally right, it should show up here first — because trading is a domain where the feedback loop closes within days, where data compounds faster than models improve, where orchestration (getting the right signal to the right decision at the right moment) is most of the work.
The 1M-context window is the specific feature that made solo feasible. Things I'd normally hand off — spec reviews, refactors, doc drift — I now do in the same session.
I'm the architect, the ops engineer, the trader, the skeptic. Claude is the pair programmer, the research assistant, the code reviewer. We are coworkers — hence the name.
The platform integrates institutional flow data, catalyst news, per-ticker Bayesian calibration, adaptive thresholds, and a weekly calibration loop that's still wrong often enough to teach me something.