ADR-065: Theories can be contested: contested_by and rivals

Context

A practice could be contested from the start: SOTA has carried contested_by: since the schema was migrated. A theory could only be corrected. corrects says an account has replaced another’s reasoning, and nothing in the THEORY scheme said that an account is disputed and nothing has replaced it yet, which is the ordinary state of a Proposed theory. The dispute lived in prose and in inactive-ok directives (“one of the three rival mechanisms this cluster holds”), where nothing could check it and no view could show it.

Two shapes of dispute were already in the record:

• By evidence. THEORY-024 was demoted to Proposed because LIT-456’s geometry-free control matched Muon. THEORY-072 and THEORY-071 both cite LIT-537’s grokking run with no weight decay as the counterexample they cannot absorb. The separator account filed on this branch, THEORY-117, has LIT-414’s small value norms at [SEP] against it.
• By a rival account. THEORY-070 was filed with corrects on THEORY-072 and THEORY-071, and its own promote_when said it “corrects the weight-norm accounts without replacing them”. That is not what corrects means. THEORY-033 says a spectral optimizer pays for its step size and not for its geometry, where THEORY-024 says the geometry is why; neither has replaced the other.

The owner asked on #397 for contestation relations on the theory schema.

Decision

Two relations on THEORY, both explain: cited (ADR-063):

• contested_by, scheme LIT, no converse: the paper whose evidence disputes the account. It mirrors SOTA’s field exactly, including the reason it has no converse: the far side would be a docket on a note.
• rivals, scheme THEORY, its own converse: another account of the same phenomenon that cannot also be right, where neither corrects the other. Symmetric, like compared_against on LIT: the fact is one fact, so luria link --fix writes the other side.

When a dispute settles, rivals becomes corrects on the winner and a status on the loser. Two accounts that might both be true are not rivals: THEORY-032 and THEORY-033 locate the spectral optimizer’s advantage in different places, and both documents say whether they are one mechanism is open, so they are left unrelated. Likewise THEORY-117 and the no-op reading it argues with are “not exclusive”, so it takes contested_by and no rivals.

Applied in this contribution:

• THEORY-070: corrects → rivals on THEORY-072 and THEORY-071, at version 2.
• THEORY-072 and THEORY-071: contested_by: [LIT-537].
• THEORY-024: contested_by: [LIT-456], rivals: [THEORY-033].
• THEORY-117: contested_by: [LIT-414].

Alternatives considered

• Use corrects for disputes too. That is what THEORY-070 did, and it made corrected_by on the two weight-norm accounts read as if they had been replaced, while the document doing the correcting said they had not. A relation that means two things is checked as if it meant one.
• One contested_by accepting either a paper or a theory. A paper’s evidence and a rival account are different kinds of dispute: one is answered by a result, the other by a mechanism that explains more. A rival is also symmetric and a paper’s evidence is not, so one field cannot declare the right converse for both.
• Status quo. The disputes stay in prose and directives, invisible to the views, and corrects keeps its second meaning.

Consequences

• CLAUDE.md’s list of relations that must be explained in the body names both.
• A rivals relation needs a sentence on both sides, because the converse is held too and is also explain: cited.
• The three grokking accounts are now visibly a dispute rather than a chain of corrections, which is what THEORY-069 says they are.