[PierreMardon]

#123 changed the blueprint the extractor draws, and the judge's method did not follow (#124).

• The frame. count.md and the no-padding criterion described the frame as the idea, the criteria and the scope, then the body and the sensitive zones. The section "Assumptions that approval confirms" was not in it: its callout and its items, which name rules the page states elsewhere, could be counted as padding.
• Pointers. count.md gave a short pointer such as "see criterion 3" as the way not to repeat. The assumptions approval confirms gathered after the opening, and a blueprint understood where each passage stands #123 asks for the opposite: a fact named in a few words where a passage needs it, and a reference that adds to a sentence, never stands in for it. Read as it was, the method counted the few words as a repeat and let the pointer through.
• Nothing measured the defect. No measure of the count saw a reference that stands in for what a sentence says.

What changes

The method

• The section of the assumptions is part of the frame, in count.md and in the rubric. Its callout, its items and the mark where a rule is stated say no rule twice and are not padding. What an item says past the name of its rule, what left it open and the alternative set aside, is a fact like any other.
• A fact named in a few words where a passage needs it is no repeat, and no pointer is given as the way out of one.
• no-invented-rule speaks of an assumption that approving the blueprint confirms, no longer of one "to confirm".

A fourth measure

• references_standing_in: the passages that, their reference taken away, no longer say what they mean. Such a passage refers by the number of a criterion, a label, a code or a title, or sends the reader to a file they do not read.
• A reference that adds to a sentence complete without it is not counted. The judge lists each one with its passage, and the script counts them as it counts the details.
• assumptions is a place a statement stands in.

The check of the judge

• Each pair names the measures of the count its spoiled page must raise. The padded page raises the three it raised, and need hold no more references.
• A new page, blueprint-referring.md, holds the facts of the lean one, with references in place of the words that name them. It must raise the references. It spoils no criterion of the rubric, so the count alone checks it.
• An original that two pairs share is counted once.
• The lean and the padded pages carry their assumption in the form of today, so the originals of the check hold the section the method now sets aside.

The report

• It folds the measures that a run counted before holds.
• It names as not compared a measure of the count that one summary alone holds. Against every campaign kept, that is references_standing_in.

The pages

• evals/README.md defines the measure and the check.
• evals/reports/README.md says what every campaign kept cannot be compared on. A move of blueprint.no-padding, repeated_in_prose or skippable_percent may come from the judge.
• ARCHITECTURE.md changes in two sentences: what the count counts, and what the check holds of it.

Checked

• scripts/gate.sh, green locally: 1,902 tests on 3.11 and on the newest Python.

Not measured: no judge check and no campaign was run. Both are billed and wait for your go. The next step of #124 is a campaign that measures #123 against 2026-10-07-*.

Closes #124