The chess habit my AI harness was missing

2026, July
A late-endgame chessboard with few pieces remaining, handwritten scoresheets resting beside it, one player's hand pointing at a square mid-conversation, lit by warm amber candlelight against near-black shadow.
AI Engineering Opinion

After a bad tournament loss, I sit down with the person who beat me and replay the critical moments.

Chess is a battle of ideas.

I might spend forty minutes trying to prove a sacrifice works while my opponent spends the same forty minutes trying to prove it doesn't.

The critical positions are where one of us runs out of answers.

Sometimes I've lost because I kept trying to prove a point after the board had already settled it.

The move you want to defend deserves the hardest review

The habit that opens this piece has a name: the post-mortem.

It's common after a tournament game, though nothing requires it: players sit down together and go through what happened.

There's a second habit, for the move you haven't played yet, the one you're proudest of.

Before you commit, you pause and look at the position through a beginner's eyes: is there a mate you missed or a piece you left hanging?

It's called Blumenfeld's rule, after the Soviet master who first described it.

The move I want to play most is usually the one I try hardest to break.

I don't need to like how feedback arrives to test whether it's right.

My opponent spent the game trying to refute my idea.

That's exactly why they saw what I couldn't.

A passing check can still miss the decision

My AI harness had reviewers, but they judged against rules alone.

None of them were trying to prove it wrong.

It could tell me my code passed every check I'd written. It couldn't tell me the check itself was wrong.

A feedback system that ends in an archive is a filing system.

Feedback earns its value when it changes the next decision.

My harness preserved corrections, but it didn't load them before the next run.

During the redesign of my AI harness, one check confirmed that a safety change existed in the code.

A second review traced the decision it controlled and found the change wired into the wrong branch.

The check passed on existence, not on the path that mattered.

I changed the standard after that: a safety rule only counts once the decision path proves the outcome it produces.

Chess taught me to stop defending a move after the board had refuted it.

The harness taught me to stop defending a system after the evidence had done the same.

Feedback is useful long before it feels good.

What does your next decision know that the last one didn't?