ME.22:4.3 - Observe the receiving use

Ask the reader to obtain the first useful result, not to rate which document looks better. Recover:

• the applicable action or decision and the answer actually given;
• the result, stop or missing condition the reader identifies;
• unsupported inferences or lost distinctions that change the answer;
• the source returns needed to support or correct it; and
• the observed burden when time, source openings, attempts or other relevant effort were actually recorded.

When the question is how much selected structure the reader can extract, use C.2.8 U.ExtractableStructuralInformation. Identify the expressed account, its publication form and the reader, with selected relations, correctness criterion, prior knowledge, operations, help, access and budget qualified for this comparison. A qualitative account of recovered and missing relations often suffices. A number needs a declared structural scale and fixed units or denominator, not counts of words or arrows.

Keep a design estimate separate from an actual recovery observation or a mapped formal estimate. A response establishes recovery in that trial, not a maximum or population reliability. State what changed in content, what changed in expression, and what preparation, actual help or a source return supplied. Familiar relations can count; a reader’s independent correction of false or absent instruction is not credited to the expression. Missing basis is not zero. A successful changed-condition return supports its own condition rather than increasing the original amount.

These are candidate observations, not one universal score. Select those that can change this comparison. A source-opening count does not measure comprehension; shorter reading does not prove lower total effort; recognizing a prescribed action does not establish its correct enactment.

Choose readers and allocation to variants in proportion to the claim. A fresh probe withholds the other answers and the author’s preferred conclusion when they would teach the task. Repeated-reader comparisons retain learning and order effects instead of calling the reader fresh. In AI probes, keep the model, supplied context and available tools recoverable when they affect inference. Generalizing from a few AI answers to human performance is another unanswered question.