How Strongly Does the Text Say It? An Evidence-Grade Protocol for Interpretive Claims in Machine-Assisted Work

Every interpretive claim — about what a text means, what a passage's structure is, what a law requires, what an instrument does, what a source supports — rests on evidence of some strength. Almost no interpretive work records that strength. It records the conclusion, and the conclusion then travels: into a summary, a citation, a review, a decision, a system's training data. By the time it arrives anywhere that matters, a careful reading supported by the text but open to a strong alternative has become a flat statement of fact, and nobody along the way decided to make it one. This article specifies a protocol for stopping that. Its central principle is Claim-Grade Separation: every interpretive claim is recorded with its evidence grade as a separate field, beside the claim rather than inside its prose, so that the grade survives reuse. Hedges in prose are the first thing compression removes; a grade in its own field is out of compression's reach. The protocol defines six grades on one ordinal scale — G5 Explicit, G4 Entailed, G3 Supported, G2 Plausible, G1 Conjectural, G0 Contradicted — with a five-question decision rule in which the decisive question is whether the strongest alternative reading has been stated and found weaker. It argues that a grade is not a probability and cannot be averaged. Three failure conditions are named. Confidence Flattening is the delivery of every claim at one register regardless of support, characteristic of generated text, which has no separate channel for the strength of its grounds. Grade Inflation is the upward drift in apparent confidence as hedges are dropped in each retelling while the evidence stays fixed. Aggregation Error is grading a conclusion above its weakest necessary premise, prevented by a weakest-link rule under which even several independent moderate confirmations do not raise a grade. The protocol has seven steps and specifies whether a machine may perform each. Machines may locate evidence, help generate alternatives, and assist with controls; they should enforce the carrying-forward of grades, which requires consistency rather than judgment. They may not assign a grade and may not stand behind one. The negative control — testing whether the pattern a reading relies on also appears in shuffled text, comparable text, or under random boundaries — is specified as the protocol's most important safeguard, with a catalogue of four forms and the requirement that passing conditions be fixed before a control is run. The protocol generalises the evidence-grade coding the author developed for structural analysis of dense scriptural and classical text, and it is the method all ten papers in The Answerability Series used. The article applies it to itself: every substantive claim in its core sections carries the author's grade in brackets. It closes by reading the ten papers as one argument — that wherever a claim is made, an answer given, or an obligation set, reliability depends on whether anyone is positioned to be asked about it — and includes a series map of every construct and instrument released. Worked examples across textual structure, regulatory obligation, source summary, and historical intent show the protocol catching a reading the text excludes and a paraphrase presenting a G2 finding as G5. An adoption ladder begins with a ten-minute step: grade your own next three claims. A table of actions by role covers scholars, editors, lawyers, analysts using AI tools, system builders, and teachers. Paper 10 of 10 — the final paper of The Answerability Series. Manuscript ID EGP-METHOD-2026-10. 38 pages, 10 figures, 21 tables, seven appendices including a grading worksheet, a grade reference card, a negative control catalogue, and the series map. Supplementary files deposited with this record: a grading worksheet (CSV) and a JSON schema that refuses to accept a grade above G2 without a stated alternative, or any grade without a named grader — all released under CC BY 4.0.