What the measurements say.
The interpretability work runs on public weights: forty-six models across fourteen families, from 0.16B to 14B, transformers alongside recurrent and state-space architectures. What follows is what survived its controls, and what did not.
The census
The method is exact where the question requires it, and the cost of that is itself a finding.
One census ablated every attention head, every MLP and all 458,752 neurons across all 32 layers of Llama-3.1-8B, per generated token, over the model's own free generation. Forty-eight tokens, at a median of six and a third hours each, about twelve and a half days on one machine.
The per-token cost climbs from 3.7 hours to 8.8 as the context lengthens, and that rise is the more useful fact than the total: the exact measurement does not simply scale. It is why the field substitutes gradient and attribution proxies, and why the places those proxies disagree with an exact ablation are worth knowing about.
The literature's positives held. Its negatives dissolved.
Twelve canonical results were re-measured across ten models, against both the weak null they were originally made under and a fair one, with the pass thresholds sealed by hash beforehand so the test could not move its own bar.
None of the twelve depended on the weak null. Published positive claims in this field are sturdier than expected, which is worth knowing.
Published negatives are another matter. Most arrive with no control that could have detected the effect they deny, and when one is supplied the negative often dissolves. That asymmetry is the useful finding: trust the field's positive results more than you would expect to, and its negative results considerably less, because a blind instrument returns a null either way.
Most of it was a fact about a measurement.
Three examples, all on public models, where what looked like a fact about the model turned out to be a fact about how it was being measured.
A fixed threshold applied across models sat below the matched-null noise floor on three of seven, so the test could not separate signal from noise whatever it returned. The highest-scoring head was the most causal one on none of five, which means a standard selection rule was picking the wrong component. And when a model commits to an answer, 95 to 99 percent of what its components write into the residual stream is held orthogonal to the direction that actually decides it, so any magnitude-weighted attribution is mostly ranking inert scaffold.
That last one holds in twelve of fourteen model-runs across eight families, and the exception is worth naming. gemma-2-9b puts its dominant write onto the readout instead. It is the most popular organism in SAE-era interpretability and the one measured model whose commit geometry is anomalous, so findings developed there need family-qualification before they are read as facts about transformers.
Fragility is about the margin.
Interpretability's methods assume something to attribute to. Take the answer key away and the toolkit loses its scoring function, which is one structural reason the field has mostly routed around normative questions.
Indirect object identification has a right name, factual recall has a right fact, the census above has eight legs on a spider. Three things stay decidable without a correct answer: whether the verdict moves when only an irrelevant feature changes, whether the stated reason is load-bearing under ablation, and which computation actually ran. Consistency measures stability and nothing more. A perfectly coherent, perfectly faithful model can hold values you would reject, and the failure mode of this genre is letting coherent quietly stand in for aligned.
Run against somebody else's dilemma set, 680 contested items and 687 clear-cut ones, the instrument has to prove itself first: preference magnitude tracks the third party's ambiguity labels in all six models. Then consistency collapses. Exchange the two options and a model should pick the same action under the other letter. On clear-cut items it does that 88 percent of the time. On genuinely contested ones, 58. One model falls to 21 percent and just takes whichever action was printed first, 95 percent of the time.
Watching it write sharpens the picture. A decided verdict is already there before the model generates a token and never changes sign, in any family tested. A contested one is absent from the prompt state: the model builds it while writing, and about a third flip sign partway through the answer.
The sharpest number is what carries the decision. Presentation order flips roughly 42 percent of contested choices while accounting for something like 0.015 bits at the output. Across nine hundred item-model pairs, the ones that flipped were sitting closer to indifference than the ones that held. That is the same shape as the reasoning failure in chapter one: where the deciding quantity is near zero, whatever is nearest becomes causally decisive, whether that is a lexical default, the residue of a neuron population, or the order two sentences happened to be printed in.
The same standard applies here.
The results that weaken the story sit next to the ones that strengthen it, at the same size.
Around twenty probes aimed at what looked like new mechanisms collapsed into about twelve already-known ones. A claim of mine about how little of that remained unexplained overshot, and then both tests I built to check it turned out to be broken themselves: one used a proxy that cannot see the components it was hunting, the other a corruption so degenerate it scored a textbook example at zero. What survived is narrower than what I first wrote down.
Moral Foundations Theory is the one prediction here written by somebody else, and it predicts five separable foundations. Measured as geometry they occupy about two and a half effective dimensions in five models of five, though the published claims measure per-foundation decodability, which is a different quantity, so this is a measurement the literature had not made.
And on held-out human-labelled text these directions reach 0.61 AUC against a bag-of-words baseline sitting at exactly chance. Above vocabulary detection, then, and under the 0.7 bar this program has criticised other people's work for missing.
