LLM Judge Training Data: My AI Checker Changed 12 of 29 Grades

0:00Anthropic trains its AI partly with feedback from another AI.

0:03But there's a problem: the AI giving the feedback can be wrong.

0:07Here's how I built an AI judge for my trading agent that I can actually trust.

0:11Recap: I'm building an AI agent that builds profitable trading strategies, starting with a critic that checks its work.

0:17The agent gets scored on real results: live or paper trading when a strategy has them, and a walk-forward test on time periods it never saw when it doesn't.

0:26And I already have over a thousand paper-traded portfolios my agent built.

0:30To train that critic, I need an answer key: real agent runs, each with the right grade.

0:36The grade is good, repair or bad.

0:38Good means the work was done right, repair means it has a fixable gap, and bad means something critical is wrong, like made-up numbers.

0:46This matters because later, the critic will decide which runs count when I train my own model.

0:51So one wrong answer in the key can spread.

0:54Here's how I built it, in two steps.

0:57Step one, I used real runs from my agent: what my users actually asked, and everything the agent actually did.

1:04Each run becomes a packet: the user's request, the agent's full instructions, every tool it called, what came back, and the answer being graded.

1:12Step two, I checked the answer key itself.

1:15My first grades came from Codex and Claude Code sessions, so I needed something to check them.

1:19First I ran a bakeoff of thirteen decision models, which are built to answer quick multiple-choice questions.

1:26Most of them answered good to everything, and the best got nineteen of twenty-nine cases where I already knew the answer.

1:31They failed because the mistakes hide across a long run. One run said a strategy paid thirty thousand dollars in fees. Seven hundred seventy-one lines later, it said the strategy never made a single trade.

1:43Then I tried DeepSeek V4.1 Flash, and it did best, with twenty-one of twenty-nine.

1:49But when I ran it three times on the same cases, it changed its grade on twelve of them.

1:54A checker that changes its mind that often can't tell you which grades are wrong.

1:58So I switched to blind review, using Claude Code subagents, which are separate AI reviewers that each start fresh.

2:05Blind means each reviewer sees the raw run and never the old grade, so it can't just agree with it.

2:10Each one writes a grade, a reason, and exact quotes with line numbers.

2:16Then I lock those verdicts, so nobody can quietly change them, and only after that do I reveal the old grade.

2:23When they disagree, a fresh reviewer checks every claim against the source.

2:27And when the written rules didn't cover a case, I made the call myself, nineteen times.

2:32In the end, seventy-one training grades changed.

2:35Some were too lenient.

2:36Others were too harsh, so twenty-seven plan and step grades went up to good.

2:39Even the blind reviewers got overruled more than a third of the time.

2:44What's left is eight hundred eighty-nine examples I trust.

2:47If you want to see how I train it, like and follow for part three.