Coherence is not truth

Put a group of people in the same workshop for long enough and they start talking in shorthand. The shorthand is useful. It is also private: a newcomer, or someone from the shop next door, hears the same word and takes it to mean something else. Nobody notices, because everyone is nodding.

Software agents do the same thing. Left to adapt to their own work, their vocabularies drift apart. The usual fix is to forbid the drift: one dictionary, one shared set of meanings, everyone speaks the same way. Our second publication tries the other approach. Let them drift, and build the machinery between them that keeps them understandable to each other. Then find out, honestly, where that machinery stops working.

A three-way check caught misunderstandings that no two-way check could see, and fixed them exactly. It also has a hard edge, and we measured where it is.

The experiment

We built a small population of agents with a made-up language, simple enough that we can inspect every word. Their vocabularies were allowed to drift. Shared words shifted meaning by topic, and pairs of words that kept turning up together got fused into new ones, the way “web log” became “blog”. Every fusion was written down, so any word’s ancestry could be traced.

Between any two agents sat a translation: not a model, just a small, readable table learned from a few examples they both agreed on. That is the whole trick. When the translation is ordinary data, you can check it, audit it and, when it goes wrong, correct it.

agent A → translate → agent B → translate → agent C → translate → back to A

Here is the useful part. Translate from A to B, then B to C, then C back to A. You should arrive back where you started. If you don’t, something in the loop is wrong, and the gap tells you exactly which words are affected and by how much.

What the loop saw

We planted a translation error that two agents could not detect between themselves. Every back-and-forth between them looked clean. But the moment a third agent joined the loop, the gap appeared, and it named the affected words one by one. Feeding that gap back into the translation repaired it completely. Not most of the way. All the way, at every seed we tried.

Two agents agreeing with each other is not evidence that either of them is right.

Where repair stops

Some damage cannot be undone, and we can say precisely which. If an agent has merged two different words into one, the information is gone and no amount of loop-checking gets it back. If it has swapped two words, that looks just as bad but is fully repairable. We tested this on hundreds of planted collisions and the rule never broke.

Where blame stops

Knowing that a loop is broken is not the same as knowing which link broke it. With only three agents, you can never tell: the loop reports a total, not a culprit. With a better-connected population, you can, and the boundary turned out to be simple and local: an accused link can be pinned down when the agents at either end have enough other connections to cross-check it. Inside that boundary, the instrument was never wrong. Outside it, left to guess, it named the wrong link most of the time. So we made it refuse instead. A translation that has real evidence behind it is never overwritten automatically, and that refusal fired every time it should have.

The result we didn’t want

We also put a price on all this. In a small simulated economy where agents earn by completing tasks and pay for the words they send, letting vocabularies drift and instrumenting the drift beat forcing everyone to speak alike, but only inside a window of bandwidth cost. Outside that window it lost, and we say so.

Inside the window, the value came from sharing definitions and repairing by ancestry. The three-way loop audit, the thing this whole note is about, earned nothing there. Switching it off changed the outcome by nothing at all, and a preregistered confirmation on fresh seeds agreed. That narrows where the instrument is worth running. It does not mean it is useless, but we would rather publish the number than the slogan.

Coherence is not truth

A final experiment, planned and frozen before we ran it, stacked three separate questions on one action: was it translated right, is the evidence right, and is the agent allowed to do it? Each can fail on its own, and fixing one never fixed another. That is the point of the title. Three agents can be in perfect agreement and all be wrong about the world. Agreement is a property of the relationship between them. Truth lives somewhere else, and an architecture should never let one stand in for the other.

That experiment’s headline number came in at a perfect score, and we are reporting it as unsupported anyway, because fewer runs qualified than we had committed to in advance. The design said to print that outcome without apology, so we have.

What we are not claiming

This is a toy world, built so that every word can be inspected. The maths behind it is standard, and the note says where it comes from. Nothing here shows that two real language models talking in English would behave this way. They share a surface, not a vocabulary of this kind, and there is no translation table between them to audit. What we have is an existence proof that divergence can be instrumented, with its limits stated as precisely as its powers.

Why we think this matters

Our first paper, Where reliability lives, argued that some of an agent system’s reliability belongs to the institution around the model rather than to the model. This note applies the same idea to the conversation between models. Drift is allowed. Misunderstanding gets a type. Some of it can be fixed automatically, some of it has to be refused, and none of it gets to impersonate the truth.

Where to read it

The note is on our research page, published here rather than as a preprint. The programme note is the shorter read with the actual figures. Alongside them are the verification packets: the search record behind every claim that something has not been done before, the fingerprints of every frozen result, and the receipts from the review panel and the full test run. Every number in those documents is copied unchanged from the private record. Reviewers who want supervised access to it can write to [email protected].

— Taniwha AI, September 2026