Gemini reasons top-down: it imposes a taxonomy, populates it with instances, and presents the taxonomy as the finding. The opening move is definitional (“the intention gap is a structural, cross-domain failure mode”) and the report structure is the argument. This is the shape of a well-produced research-firm deliverable: clean, orderly, hard to argue with sentence-by-sentence, and optimized for the reader who needs a framework they can immediately use.

Claude reasons bottom-up: it opens by undermining Maoz’s own framing (“Maoz’s argument should be read, against its own framing, as a symptom rather than a warfare-specific claim”), and spends the document re-scoping the question to what the primary-source evidence can actually support. It pushes back on interdisciplinarity as a catch-all fix, reformulates it as Galison’s specific “trading zone” institutional form, and flags where Maoz’s own empirical neuroscience cuts against his op-ed’s policy claim. This is the shape of an essay, not a deliverable.

Neither approach is wrong. They are the signatures of two different inductive biases, built into the models by the training choices of two different labs. The question the rest of this dispatch asks is whether those biases are architecturally determined— which is to say, whether they follow from the physics-shaped information-theoretic framing Anthropic inherited from Kaplan’s scaling-law work, and from whatever the analogous deep prior is inside Google DeepMind.

A side-by-side failure analysis is where the divergence stops being aesthetic and starts being structural.

The citation count gap — 43 sources vs. 2,501 sources consulted — is not a quantity-vs.-quality argument. It is evidence of two different conceptions of what a source is for.

Gemini’s 43 sources are bibliographic: they anchor the report’s legitimacy. They appear at the end. Their job is to prove the framework rests on something. The reader is not expected to follow any given citation; the reader is expected to be reassuredby the list’s weight.

Claude’s 2,501 sources are argumentative: they are the load-bearing substance of individual sentences. Specific arXiv IDs, specific papers by specific authors, quoted or named at the point of use. The inline citation chips (“[LessWrong]”, “[arXiv]”, “[Transformer Circuits]”) are there because the claim would not otherwise be defensible. A reader who drops any single source finds that paragraph wobble.

Everything above is texture. Here is the demonstration. One sentence from each report on the same underlying question: what is the mechanism that makes the interpretability gap persist?

The Gemini sentence is true. It is also the kind of sentence that has been true for the last seventy years of strategic-studies writing about arms races. It is framework-level.

The Claude sentence is specific. It names the paper, the three comparative statics, and the non-obvious one — that more information makes race dynamics worse, because leaders use transparency to shave precaution to the minimum that preserves their lead. This is the operational claim. It tells you what to do: verification regimes that reveal compliance without revealing exploitable capability.

One is the headline. One is the mechanism. They come from the same prompt. The difference is not in what the models know. It is in what they are built to do with what they know.

The single most important piece of news in Claude’s report does not appear in Gemini’s. From the Claude run:

This is sourced to Neel Nanda’s team’s “Pragmatic Vision for Interpretability” post on the Alignment Forum, which quotes an internal admission that “SAEs underperform much simpler & cheaper techniques.” Three frontier labs are visibly making different bets on the same tool at the same moment: Anthropic is still on SAEs-plus-attribution-graphs, OpenAI used Goodfire’s SAEs rather than its own for its 2025 misalignment work, and DeepMind has quietly walked away from SAEs as a primary research direction. The popular narrative of monotonic interpretability progress obscures what is actually a paradigm split with real technical disagreement.

Gemini’s report covers sparse autoencoders cleanly — it cites Gemma Scope 2 as DeepMind’s contribution, frames SAEs as “the most significant technical breakthrough of the 2024–2025 period,” and moves on. It is factually accurate. It is also the view that would be accurate if DeepMind were still committed to the program. The report misses the pivot.

Why does Claude catch it and Gemini miss it? One honest answer: Claude is produced by a lab where interpretability is the founding identity, and its research mode is tuned to surface internal methodological disputes within the interpretability field. Gemini’s research mode is tuned to produce clean summaries of what the literature says, not to weight internal methodological disputes over the surface consensus. On a meta-interpretability prompt, one of those tunings catches the news and the other produces the literature review. That is what architectural determinism means when it leaves the whiteboard.

Three takeaways, ordered by how much they should change your behavior as a serious user of these tools.