Criterion V3 — Alternative Level

Validity type	Interpretive
Pass condition	The evidence is shown not to be equally well explained by a claim at a simpler description mode
Evidence family	N/A (criterion is about claim structure)
Minimum reporting	The description mode claimed, the next-simpler mode considered, and what the evidence rules out at that simpler mode
Common failure mode	Claiming an algorithmic-level finding when the evidence demonstrates only an implementation pattern

What this criterion requires

Alternative level asks whether a claim pitched at one description mode could be fully accounted for by evidence at a simpler mode. Craver (2006) distinguishes “how-possibly” explanations, which show a mechanism could produce a phenomenon, from “how-actually” explanations, which show it does. Evidence sufficient for a how-possibly account at a lower mode is routinely reported as though it had settled a how-actually claim at a higher one — the gap between the two is where alternative-level failures live.

Satisfied when:

1. The claimed mode is stated. The study names the description mode the claim occupies (see the seven description modes, which extend Marr’s three levels).
2. The next-simpler mode is considered. The study asks whether the same evidence is equally consistent with a claim one mode down.
3. The simpler mode is ruled out, or the claim is downgraded. If nothing in the evidence distinguishes the two, the stronger claim is unsupported and the verdict is stated at the lower mode.

MI example

Othello board-state probing recovers linear representations of game state from network activations, and causal patching on those representations shifts model predictions. That evidence licenses a representational claim: a causally-used state summary. The published label “world model” asserts something a mode higher — a process resembling planning or simulation over that state. Nothing in the probing or patching results distinguishes a world model from a causally-used summary that a simpler mechanism could produce equally well; the gap runs between evidence for the first reading and a name that claims the second.

Relation to other interpretive criteria

Alternative level (V3) presupposes level declaration (V1) and level–evidence match (V2): a claim must state its mode and have evidence sufficient for that mode before it makes sense to ask whether a lower mode would do as well. V3 is the parsimony check that runs once V1 and V2 are satisfied, and it feeds directly into unlicensed labeling (V4) — a claim that defaults to the higher mode without running this check usually does so through a name that already presumes it.