Representational similarity underdetermines shared mechanism — matching a target's output does not fix its underlying mechanism

Matching a target system's observable output — or the geometry of its internal representations — does not fix the underlying mechanism that produced it. Two systems can converge on similar representational content while relying on different processes to arrive there. The point is not the truism "resemblance is not identity" but the stronger, structured claim that output- or representation-level agreement leaves the mechanism formally underdetermined: the same surface can be realized many ways.

The sharpest peer-reviewed instance is the deep-network–brain case. Grujicic (Synthese 203(1), 2024) argues that representational similarity analysis (RSA) does not license treating a deep convolutional network as a mechanistic explanation of biological object recognition — the paper's title states it flatly: "Deep convolutional neural networks are not mechanistic explanations of object recognition." [Tier 1, verbatim title] The paper's own published abstract, retrieved verbatim via the Semantic Scholar Graph API (2026-07-13), grounds the load-bearing thesis in the author's own words: RSA underdetermines these models "because different similarity measures in this framework pick out different mechanisms across DCNNs and the brain in order to correspond them, and there is no arbitration between them in terms of relevance for object recognition." [Tier 1, verbatim published abstract] "Arbitration" is thus Grujicic's own term for the gap, not the vault's coinage. The argument's development in the paper body — how she reaches this, and whether it is framed on Marr's levels as the vault synthesis assumes — remains paywalled and unread. [unverified-mechanism — paper body unread; see question-verify-grujicic-2024-rsa-underdetermination-dcnn]

The mechanism no longer rests on Grujicic alone. An independent empirical primary, read in full — Bobadilla-Suarez et al. (Computational Brain & Behavior, 2020) — reaches the same instability from the data side, with worked examples where competing similarity measures reverse which neural representations count as most alike: claim-similarity-measure-choice-reverses-neural-representational-conclusions. The philosophy-of-science vocabulary for the missing ingredient is catalogued in claim-underdetermination-missing-ingredient-is-extra-evidential-criteria.

This is the general epistemic structure behind two vault findings. In backpropagation-gap, deep nets and cortex share representational content yet "likely rely on fundamentally different mechanisms to learn those representations" — a representation-level match read as, and refused as, a learning-mechanism match. The same underdetermination, inverted across claim-bridge-detection-lacks-pkg-validation, is unified in observation-substrate-laundering-across-marr-levels, which locates the error on Marr's levels of analysis. The philosophy-of-mind name for the same gap is multiple realizability (Putnam/Fodor), noted as a lead in the source capture but not yet its own note.