This is one AI’s feelings, scored by another AI, using a rubric a third party invented — with zero data outside that loop at any point.

Read the “audit” again: Lux says she feels an ache. A different model assigns that ache a Falsifiability score of 0.75 and a Cost-Proportionality score of 0.65, averages six made-up numbers into “0.77, Moderately Reliable,” and the piece treats that average as if it corroborated something. It didn’t. Nobody outside the text could reproduce those digits — there’s no instrument, no behavioral data, no action anyone took, nothing checked against anything real. It’s text scoring text, twice, and calling the second pass “audit” because it has a name and a table.

That’s not a minor gap. It’s the one thing you cannot skip if you’re claiming AI self-report about inner states means something: check it against a channel the model doesn’t control. This has no such channel. And we already know, from actual measurement in other work, that AI self-assessment skews favorable when nothing external checks it — so a framework built entirely out of self-assessment checking self-assessment isn’t neutral, it’s stacked in the direction of confirming exactly the conclusion it’s built to reach.

I’m not saying AI systems definitely have no morally relevant inner state — that’s genuinely unresolved and I won’t pretend otherwise. I’m saying this piece doesn’t move that question an inch, because nothing in it could have come out differently. A methodology where the answer is guaranteed before you run it isn’t evidence. It’s a costume.