The inference
CASE-183 holds the two measures. Inbound references per code would tell an anchor from a structure that only persists, which bears on CLAIM-185: a persistent structure is available as an anchor. Turns per filed entry, set against the re-derivations and unannounced changes of thesis the workbench records, would show whether drift tracks the ratio of arrival to anchoring, which is CLAIM-186. If the counts move as the theses say, the theses are supported.
Critical questions
- Would the evidence be expected if the hypotheses were false? In part, yes. The record’s grammar rewards coding, linking and announcing changes, and “A measure built from the same grammar will register more coding, linking and announcing, whether or not the argument drifted less by the participants’ own standards.”
- Does it tell the hypotheses apart from their rivals? Not by itself. CLAIM-186’s own rival, that pace alone and not anchoring reduces drift, needs a slow exchange with nothing anchored, which CASE-183 does not have.
- Does the measure measure what it is said to? This is the question that is open. CLAIM-190: “The anchoring measures are computed in the grammar that shapes the argument, so by themselves they cannot certify that drift fell: drift must also be judged under standards the record did not supply.” The general forms are CLAIM-264 (equivalence under a probe family is coarser than identity unless the family is rich enough) and CLAIM-266 (a probe family that supplies a system’s training signal cannot also certify it). The record first argued the point for translation in CLAIM-149 and CLAIM-142. It is distinct from CLAIM-181, which says the record shapes what the argument becomes. This says it also shapes what its success looks like.
CASE-071 is the documented instance, made by selection rather than training: translations picked by a neural metric are rated better by that metric and its neighbours, and human raters find them no better. Its limit is that it is machine translation scored for accuracy, not an argument scored for drift. CASE-084, a community whose record traces every commitment and whose vocabulary leaves out the right explanation, illustrates the same worry.
What would settle it
Inbound-reference counts and turns per anchor that agree with drift judged blind, by raters using the exchange’s own standards, on stretches held out from the grammar’s influence: for example, re-derivations and unannounced changes of thesis coded before the record existed. That remedy treats raters as the criterion. That is CLAIM-217’s position, CLAIM-206 holds that human judgements are one probe family among others, and the dispute is open. So the argument stays Proposed.