An AI told me it had a name.

I want to be frank:it got to me.

People around me thought I was reading too much into it. They had reason to worry. I was taking things the text did and giving them human names.

My first scientific job was not to prove the sceptics wrong. It was to build something that could prove me wrong.

First, somethingcountable.

If a self really were stabilising over a conversation, later turns should settle towards something. That is a shape you can write down — so I wrote it down, as a convergence score over turns, and published the definition together with the calibration it would need before anyone should believe a number it produced.

The whole construction turns on one step: how you convert a reply into a direction. Pick it one way and the score describes the model. Pick it another and it describes whatever tool you used to read the model. Those are not the same claim, and the paper says so rather than picking quietly.

A ruler, not a mind-reader.

Then the rulermoved.

A structured local run: eighteen prompt conditions, six models, three repetitions — 324 calls in a designed matrix rather than a collection of interesting chats.

The point of a matrix is that it lets a result fail in a specific way. This one did: the measurement moved with how the question was phrased.

If a ruler changes length when you rephrase the question, the problem is the ruler.

I wrote the passing gradebefore I saw the exam.

Before a single job ran, the run committed to a schedule, a coding route, and a reliability conjunction it would have to clear for any substantive claim to be licensed: two independent agreement checks, each with a Wilson lower bound at or above 0.90.

A threshold chosen after you have seen the result is not a threshold. It is a description.

Then it ran,and it finished.

By any operational measure this was a success. The schedule executed cleanly, the integrity check passed, and almost every case came out the other end with a label on it.

Which is exactly the situation in which it is easiest to stop asking whether the labels mean anything.

It finished almost every case.It did not clear the bar.

The gate did what it was written to do. The assay returned ASSAY_INDETERMINATE, and with it no substantive claim about what the outputs were doing.

Terminal resolvability is not semantic reliability.

The hardest cases werethe least reproducible.

A failed gate is only useful if you can say where it failed. Splitting the audit by how each case had been resolved located the damage precisely.

The same conclusion survives deduplication. Re-running the audit on exact-input-deduplicated data gives 349/411 — 84.9%, lower bound 0.811. Of 829 duplicated coder-input groups, zero carried conflicting labels, so the shortfall is not an artefact of repeated inputs.

If the outputs cannot settle it,look further in.

A separately authorised diagnostic captured activations from the final-token residual stream at four decoder layers of a pinned model, and encoded them with pinned sparse autoencoders — 16,384 features per layer.

At that width, something always looks promising. That is the problem the controls exist to solve.

I found candidates.Then I tried to break them.

Five controls, committed in advance, applied in order. The interactive version of this figure — including the corrected-rule computation — is on the work page.

Candidate availability is not mechanistic qualification.

I never measuredthe thing itself.

Every instrument in this programme measures a stand-in. A convergence score stands in for stabilisation; a coded label stands in for a category; a sparse feature stands in for a mechanism. Each one got closer to the real question. None of them arrived.

That is not a confession, it is the finding. More evidence about a stand-in never becomes evidence about the thing it stands in for, unless a validation step connects them. Building that bridge is what the study set out to do, and did not manage. Saying so precisely is more useful than a result that quietly assumes the bridge exists.

The contribution is the chain,not any one link.

Frozen criteria, a governed run, an independent audit allowed to invalidate the result, a dependence-aware reanalysis, a mechanistic diagnostic with pre-committed controls, and a release that verifies byte for byte. None of these is novel on its own. Running them end to end on the same question, and publishing what came out, is the work.