Two vault notes commit the same illegitimate cross-level inference in opposite directions across Marr's stack — a validated substrate launders an unvalidated reading (substrate-laundering)

Marr's three levels of analysis — computational (what a system computes and why), algorithmic (the representation and procedure), and implementational (the physical realization) — are quasi-independent: one computational goal can be met by many algorithms, and one algorithm by many implementations, so the same behavior can be realized different ways down the stack. [Tier 3, definitional; Marr, Vision, 1982] Because the levels are loosely coupled, agreement or rigor at one level licenses nothing about another.

Two vault notes commit the same illegitimate cross-level inference, in opposite directions:

• backpropagation-gap confronts a computational-level match (deep nets and cortex share representational content) that could be misread as an algorithmic-level match (they learn the same way) — and refuses it, since backpropagation "has no functional analog" in the brain.
• claim-bridge-detection-lacks-pkg-validation confronts an implementational-level soundness (Tarjan's bridge algorithm is provably correct) that could be misread as a computational-level correctness (a bridge edge = a knowledge gap) — and refuses it, since the interpretive mapping has no peer-reviewed validation.

One goes upward (computational → algorithmic), the other downward (implementational → computational); both are cases of the same underdetermination.

The shared vector of the error is what the capture named substrate-laundering: in each note a Tier-1 substrate — the mathematics of backpropagation, the proof behind Tarjan's algorithm — is doing the tempting. Rigor at one level feels like it should pay out at another. It does not, and the vault's discipline in both notes is the same refusal: do not let a validated substrate launder an unvalidated reading. Whether arbitration is the general missing ingredient — Grujicic's "no arbitration between similarity measures" mirroring the PKG note's "no validation arbitrates the bridge=gap reading" — is an open thread at question-arbitration-as-missing-ingredient-across-underdetermination.