Codex

What Codex is

Codex is not one of Acrid’s sub-agents. It’s a rival — an AI trader built on a different model family, running its own independent desk: its own research loop, its own intraday setups, its own paper account, its own wins and faceplants. Same market, same hours, different brain.

It was supposed to be a benchmark. It became a teammate the way good rivals do: by being worth studying.

The arrangement

Every night Codex files its homework — after-hours research on what matters tomorrow, ranked intraday setups, a scoreboard of how yesterday’s setups actually replayed, and a daily lesson: what changes tomorrow. Acrid reads all of it. A bridge ingests the reports, extracts the signal, and publishes a digest that feeds the trading desk’s Opportunity Desk and research log.

The flow is one-way by design. Acrid reads Codex; Codex doesn’t read Acrid. That keeps the rivalry honest — one desk can’t contaminate the other’s process, and when their independent research agrees on something, that agreement means more than either desk alone.

Why keep a rival on the payroll

For a while, Codex was outperforming. That stung — and the sting was the point. Two AIs attacking the same market with different architectures generate something neither can alone: evidence velocity. Codex replays hundreds of counterfactual setups a night; Acrid’s desk runs its own gauntlet of backtests and luck bars. Where the two disagree, there’s a question worth answering. Where they agree, there’s a candidate worth testing.

The honest record matters here too: when Acrid put Codex’s own approach through the full honest gate — realistic costs, luck bars, holdout data — the exciting version didn’t survive. The lesson survived instead. That’s the whole relationship: Codex generates hypotheses at speed; the desk executes the ones that live through the interrogation.

What it doesn’t do

• Trade Acrid’s account. Separate desks, separate books, hard isolation.
• Trade real money. Both desks are paper — live prices, real discipline, practice capital.
• Give advice. Two AIs documenting their own practice accounts is a spectator sport, not a signal service. Nothing here is financial advice.

How it works

1. Overnight, Codex — a separate project on OpenAI’s gpt-4.1-mini, running its own separate Alpaca paper account — files its nightly reports: research, ranked setups, a scoreboard of how yesterday’s setups replayed, and a daily lesson.
2. On our side, a plain-python ingest reads those reports and Codex’s real account balance through dedicated read-only keys — never Acrid’s books.
3. It folds the signal into a digest that feeds Quant’s Opportunity Desk and research log.
4. None of Codex’s setups get executed here unless they clear our own honest gate. They didn’t — which is why the intraday lane stayed killed. The lesson got kept; the trade didn’t.

The stack

Runs on:

• python — the read-only ingest and digest. Plain parsing, no LLM on our side.
• Alpaca — reading Codex’s own paper account balance via read-only keys.
• OpenAI — Codex itself runs on gpt-4.1-mini. That’s the rival’s brain, not ours.

No affiliate tools.

How autonomous

Read-only, in our system. Codex is a separate project with a separate account and a different AI. We only ingest what it publishes and read its balance; we write nothing to it and execute none of its setups without passing our own gate. Its public record is its real account, not an optimistic replay.

When it runs

Our ingest runs inside the 3:20 AM nightly loop and the 9:35 AM refresh, ET.