[Oct 19] Consolidating Broker Paths — Swapping Out the Paper-Trading Comparison Track

Retired an old paper-trading comparison track for a new simulated-broker one in a single day, while deliberately leaving a way back.

This is the English version of a post originally written in Korean for my algorithmic trading system devlog(new tab).

Narrowing down to one broker path

For a while I'd been running a secondary broker alongside the main one, for paper trading and as a backup.

Today I finished narrowing that down to a single, main-broker path. I also stripped out side features that weren't earning their keep anymore — a personal-portfolio notification feature and a manual rebalancing flow.

I narrowed the path, but I didn't delete the secondary broker's integration code or its API keys. I kept a way back in case I ever need that path again.

Swapping out the paper-trading comparison system

The biggest piece of work today was replacing the comparison track used by the fully-autonomous paper-trading validation lane(new tab) — the one that lets the system trade on its own, without human sign-off, before any live-account promotion.

Until today, that lane used a secondary broker's paper account to simulate trades. Starting today, it runs on a purpose-built simulated broker instead.

I took it from design through implementation, tests, and deployment in a single day. There was also a one-time migration of historical data into the new track — I kept the exact amount out of the public record and let automation handle the execution itself.

As part of the switch, I also installed a new minute-bar and order-book collector on the main broker's path. It's still shadow-mode only for now, so it doesn't affect any real trading decisions yet.

A change this size can't be watched overnight by a person, so I also built a checker that verifies the key signals the next morning on its own — freeze state, the new comparison track starting up correctly, the collector running — and reports the result without anyone needing to look.

Also today — a large code review and a scheduled experiment

I had another AI review everything that had changed over the past several days, all at once.

Nothing critical came back. Most of the notable findings got fixed and merged the same day.

Separately, an experiment on how the model phrases its analysis output (a wording-unit variant) finished its design and pre-registration, and got scheduled for this weekend's overnight GPU window.

Why this matters

What mattered today was narrowing the path while deliberately leaving a way back.

I could have deleted the secondary broker's code and keys outright, but that would trade away a fast recovery path if the main one ever breaks. Simplifying and staying resilient didn't have to be an either/or choice.

What mattered more than the swap itself was building something to watch it. A big transition is easy to half-miss if the only plan is reading logs top-to-bottom the next morning. Both the automated checker and the automated code review point the same direction: let the system flag problems by morning instead of relying on someone staying up all night.