back to the lobby

Coherence Collapse Analysis

When do extra checks stop helping?

Stacking up checks to catch a lie only works if the checks are truly independent. If they secretly copy each other, adding more does nothing. Coherence Collapse Analysis is the simple piece of math that measures the difference. The full version, with every proof, is in the paper.

The math page

Checked by proof software

Five checks that all copy each other are really one check.

Say five people check an answer and all five agree. That feels safe. But if all five learned it from the same place, their agreement is not five opinions. It is one opinion repeated five times. The number of checks looks like five. The real number is one.

This matters for AI. An AI system can run many checks on its own reasoning and still be fooled, if those checks share the same blind spot. Coherence Collapse Analysis is how CIRIS tells the difference between real checks and echoes.

Counting the checks you really have.

There is a single short formula at the heart of it. It comes from survey statistics, where it is called the Kish design effect, and the same shape shows up in voting theory and safety engineering. CIRIS applies it to AI alignment.

real checks = checks / (1 + copying × (checks − 1))

“Checks” is how many checks you ran. “Copying” is how much they overlap, from 0 (all independent) to 1 (all the same). The result is how many checks you really have.

No copying: ten checks count as ten. Every check earns its place.

Total copying: ten checks count as one. It does not matter how many you add.

This was measured, not just argued.

CIRIS measured the real-check count on its own live agent traffic, across thousands of recorded decisions. On healthy traffic it has run in the range of about seven to nine genuinely independent checks. That measurement, and how it was done, is the Constrained Reasoning Chains study. You can watch the live numbers on the research page.

What this math cannot do.

The paper proves a hard limit on itself. Some kinds of harm do not come from one dishonest part. They come from honest parts that add up to a bad result, and the paper proves that some of that harm is invisible to any checker that only looks at the parts, no matter how good. CIRIS says so plainly rather than pretending the math catches everything.

What the system does promise is the record. Every decision leaves a signed receipt that cannot be quietly rewritten, so a lie has to stay consistent with thousands of receipts it can never erase. Reading those receipts well is open work, and we say so. The books stay true either way.

The math is one part of a larger whole.

This page is the measurement. The Coherence Ratchet is how the measurement is put to work. The Federation is how it becomes something many systems share. And the full proofs, written so a computer can check them line by line, are in the paper and the RATCHET repository.

CIRISsafe by structure · open by principle · kind by design