One answer, three recall scores

Fix the question and conditions

The fictional prompt is: Where does the copper token go, and when? The fixed reference answer is: The copper token goes in the west drawer before the museum opens. The fixture set contains 12 attempts. Eight retain the four named facts while changing case, punctuation, spacing, order or surrounding wording. Four omit or change exactly one named fact.

Every rule sees the same attempt text. There is no training pass, random ordering, human rating or hidden model call. Running the published operations again on the same rows produces the same classifications. The raw JSON records the protocol and rows; the CSV presents those same rows for spreadsheet analysis.

Define the three scoring rules

Literal match requires the attempt to equal the reference character for character. Normalized match lowercases the text, removes ASCII punctuation, trims its ends and collapses repeated whitespace before comparing it with the similarly normalized reference.

The checklist looks for four published facts: copper; token; both west and drawer; and all of before, museum and opens. A full checklist match requires all four groups. This rule is deliberately simple and deterministic. It does not infer synonyms or understand negation.

Read the measured output

Literal match classified 1 of 12 fixtures as a full match. Normalization raised that count to 4 by absorbing case, punctuation and spacing changes. The checklist classified 8 as full matches: the eight fixtures authored to retain all four term groups.

The four altered or missing-fact fixtures each produced three of four checklist facts and no full checklist match. That is an internal consistency check on this fixture set, not evidence that the checklist can grade arbitrary prose.

Use the result without turning it into a learning claim

The supported result is narrow: for this fixed dataset, full-match counts changed from 1 to 4 to 8 as the rule changed. A recall report should therefore name the rule behind its number. A count without that definition hides part of the measurement.

The checklist can be useful for a personal comparison when its facts are defined in advance and the original attempt remains visible. It is not an automatic correctness test. In the Recall Deck, the reader still compares an attempt with a checked reference and chooses the label.

Limits and failure cases

These 12 attempts were authored to exercise the rules. They do not estimate how often people use each wording, and they cannot show whether anyone remembered, retained or learned the museum rule. One fictional answer is not a representative sample of school, workplace or clinical assessment.

The checklist can be fooled. The copper token is not in the west drawer before the museum opens contains every required term and would receive a full checklist match despite the negation. A synonym such as western compartment could preserve an intended idea yet fail. Those cases show why a transparent narrow rule must not be described as language understanding.

To adapt the protocol, publish the reference, fixtures and fact groups before comparing totals. Keep changed protocols under a new version, preserve earlier raw files and describe corrections. For a personal practice workflow, continue with Compare after the reveal or the publication profile.

Related guides

• Write a prompt you can actually check
• Compare after the reveal
• Read a practice record without overselling it