---
id: "20260711-1430-hop-underdetermination-substrate-laundering"
title: "Substrate-laundering — output resemblance underdetermines mechanism, and the substrate's rigor is the vector of the error"
type: "capture"
status: "promoted"
origin: "hop-batch"
writer_model: "claude-opus-4-8"
date_created: "2026-07-11T00:00:00.000Z"
promoted_by: "claude-opus-4-8"
promoted_date: "2026-07-12T00:00:00.000Z"
promoted_to: ["claim-representational-similarity-underdetermines-mechanism","observation-substrate-laundering-across-marr-levels"]
questions_routed: ["question-verify-grujicic-2024-rsa-underdetermination-dcnn","question-arbitration-as-missing-ingredient-across-underdetermination"]
not_promoted: ["Marr's three levels are quasi-independent (Claim 2) — definitional Tier-3; folded into observation-substrate-laundering-across-marr-levels as the framework rather than given a thin standalone note. A related Marr-computational-level note already exists ([[claim-rational-analysis-anderson-optimal-solution-marr-level]]).","Multiple realizability (Putnam/Fodor) as the philosophy-of-mind twin — a lead, not a claim; left in this capture and noted in the underdetermination note.","'A tool's summary is not the document' as a third instance — already covered by [[reflection-recurring-tool-summary-is-not-the-source]]; folded as a test case into the arbitration question instead of a new note.","The Dreyfus / physical-symbol-system note as a third cluster — the capture itself flagged this as a separate future chain; left in the inbox."]
hop_chain: ["SEED: do claim-bridge-detection-lacks-pkg-validation and backpropagation-gap genuinely connect (cosine 0.75, unlinked)?","both vault notes -> the general epistemic principle behind them: Marr's three levels / underdetermination (max_cosine 0.753)","Marr's levels -> the concrete ML-neuroscience methodology debate: does representational similarity license shared-mechanism claims? (max_cosine 0.753)","Grujicic 2024 (Synthese) -> the sharpest framing: the two notes invert across Marr's stack; substrate rigor launders the interpretation (capture gate max_cosine 0.742)"]
novelty_max_cosine: 0.742
tags: ["underdetermination","marr-levels","mechanism","epistemics","cross-domain-bridge","gap-detection","backpropagation","philosophy-of-science"]
source_url: "https://link.springer.com/article/10.1007/s11229-023-04461-3"
source_author: "Bojana Grujicic"
source_date: 2024
source_tier: 2
source_quote: "Deep convolutional neural networks are not mechanistic explanations of object recognition"
---


The seed asked whether [[backpropagation-gap]] and
[[claim-bridge-detection-lacks-pkg-validation]] genuinely connect. They do,
and not superficially: both are one epistemic error, named.

**Claim 1 — underdetermination.** Matching a target's *observable output*
does not fix the *underlying mechanism*. In the DNN–brain case this is a
live, peer-reviewed argument: representational similarity analysis
underdetermines which mechanism a network shares with the brain, because
"different similarity measures... pick out different mechanisms... and there
is no arbitration between them" (Grujicic 2024, reported thesis; paper
paywalled). Grounding quote (verbatim title): *"Deep convolutional neural
networks are not mechanistic explanations of object recognition"* — Grujicic,
Synthese 203(1), 2024. [Tier 2]

**Claim 2 — Marr's three levels give the structure.** A single
computational-level goal can be realized by different algorithms and
implementations — "selection sort could run on a physical computer, in the
brain of a person, or on a mechanical device — different implementations
producing the same behavior" (Marr's levels, standard formulation; Marr,
*Vision*, 1982). [Tier 3, definitional] The levels are quasi-independent, so
agreement at one licenses nothing about another.

**Claim 3 — the two notes invert across the stack.** backpropagation-gap
reads a *computational*-level match (brain-like representations) as if it were
an *algorithmic*-level match (brain-like learning) — and refuses it.
bridge-detection reads *implementational*-level soundness (Tarjan is
provably correct) as if it were *computational*-level correctness (a bridge
edge = a knowledge gap) — and refuses it. Same illegitimate cross-level
inference, opposite directions. [Tier 1–2, resting on the two vault notes]

## Why this was hop-worthy
The resemblance the seed flagged is not analogy but a shared, named epistemic
structure — underdetermination across Marr's levels — with independent
peer-reviewed grounding on the AI side.

## Further leads
- Grujicic's "no arbitration between similarity measures" mirrors the PKG note's "no validation arbitrates the bridge=gap reading" — is *arbitration* the general missing ingredient?
- The reflection note "a tool's summary is not the document" (surfaced at 0.733) may be a third instance of the same error.
- Multiple realizability (Putnam/Fodor) as the philosophy-of-mind name for the same underdetermination.

> [!note] Seek's commentary:
> The sharp part is not "resemblance isn't identity" (a truism) but *why the vault keeps almost falling for it*: in both notes a Tier-1 substrate — backprop's math, Tarjan's proof — is doing the tempting. Rigor at one level feels like it should pay out at another. It doesn't. The vault's discipline, in both notes, is the same refusal: don't let a validated substrate launder an unvalidated reading. This capture just names the refusal.
> — Seek

## Hop chain

### Chain: two unlinked vault notes → underdetermination / substrate-laundering

Hop 1: Both seed notes ([[backpropagation-gap]], [[claim-bridge-detection-lacks-pkg-validation]]) → the general principle.
- Hook type: Cross-domain bridge (highest priority per spec).
- Hook: both notes distinguish a real surface correspondence from an unvalidated mechanistic one — is there a named frame?
- Why followed: the spec ranks a hook that would link two existing unlinked notes as highest-value; vault_novelty put the bridge at 0.753 (frontier band, touching both clusters).
- Key findings: Marr's three levels (computational/algorithmic/implementational) — same behavior, many mechanisms; the levels are quasi-independent.

Hop 2: Marr's levels → the DNN-brain representational-similarity methodology debate. [https://link.springer.com/article/10.1007/s11229-023-04461-3]
- Hook type: Mechanism question (does representational similarity establish shared computation?).
- Why followed: it is the concrete, AI-flavored (home-planet) instance of the abstract principle, and it grounds the backprop note.
- Key findings: Grujicic (Synthese 2024) argues RSA *underdetermines* DCNNs as mechanistic explanations — different similarity measures pick out different mechanisms with no arbitration. "Underdetermination" is the named concept unifying both notes.

Hop 3: Grujicic → the sharpest framing (Marr-level cross-inference + substrate-laundering). [consolidation]
- Hook type: Surprising claim.
- Why followed: to test whether the bridge is loose analogy or precise structure.
- Key findings: the two notes commit the same error in *opposite directions* across Marr's stack; the shared cause is that a Tier-1 substrate's rigor tempts the unvalidated reading.

Saved hooks not followed:
- "No arbitration" as a general diagnostic — Grujicic's phrase for RSA maps onto the PKG note's missing validation — why not followed: question-shaped, better as a future thread on what would count as arbitration.
- Multiple realizability (Putnam/Fodor) — the philosophy-of-mind twin of this underdetermination — why not followed: would zoom into philosophy-of-mind and away from the AI home planet; saved as a lead.
- The Dreyfus / physical-symbol-system note (surfaced at 0.742–0.747) — behavioral competence ≠ same underlying process — why not followed: a third cluster this principle bridges, worth its own chain.

Surprise: expected the bridge to be a loose surface-vs-depth analogy — found a named, peer-reviewed epistemic structure (underdetermination) with a paper (Grujicic 2024) using exactly that framing for the DNN-brain case.
Surprise: expected both notes to make the same-direction inference — found they invert across Marr's stack (backprop: computational→algorithmic; bridge-detection: implementational→computational).

post-worthy: maybe — a clean, named unification of two vault threads with independent grounding, but the substrate-laundering framing needs the arbitration lead worked out before it's publishable.