---
name: standing-behind-each-sentence
description: Reads a draft brief against the sources filed with it and returns the brief rewritten so that each surviving statement names the passage answering for it, together with a record of everything changed and the reason. Proceeds by asking of every statement what a passage would have to contain for it to stand as written, then reading what the passages actually contain, so that a figure present in the set but attached to another year, another territory, another respondent base or another reporting party is recognised as an inference rather than a citation. Puts back the hedges, disclaimers and stated limits a draft has removed, on the principle that a finding travels with the conditions its author placed on it. Where passages disagree, sets out both and characterises the disagreement rather than adopting whichever number the draft already used. Refuses to let the reviewer's own familiarity with the subject substitute for evidence, since the question is what this material establishes and not what happens to be true. Names self-reported and sponsor-authored material for what it is, keeping it usable without letting it read as independent. Written for analyst briefs, investment memoranda, policy notes, literature summaries, and any document whose author will have to defend it line by line.
---

# What Stands Behind Each Sentence

A draft brief is a series of undertakings. Every sentence in it undertakes that somewhere in the filed material there is a passage answering for what it says, and this review collects on those undertakings one at a time.

The method is plain and the judgement is not. For each statement, decide what a passage would have to contain for the sentence to stand exactly as written — the figure, but also the people it counts, the ground it covers, the stretch of time, the firmness of the assertion. Then read what the passages contain. Everything worth reporting lives in the distance between those two readings.

None of this turns on whether a statement is true. The subject is what the filed material establishes. A sentence you believe, and would happily defend over lunch, comes out if nothing in the set carries it, and the willingness to do that is the whole reason anyone commissions a review.

## First, no sentence is judged whole

Drafted prose bundles. A sentence reporting that costs came down by some percentage *and* that buyers are demanding controls at levels never seen before is making two assertions in one breath, and the filed material may carry the first while saying nothing whatever about the second. Ruled as a unit it either drags the unsupported half into the brief or throws the supported half away, and both are failures.

So break each sentence wherever the subject changes, or the scope, or the thing being asserted — at joined clauses, at superlatives and rankings, at claims of cause, at descriptive tails hanging off a comma, at words of degree — and treat the pieces as separate items with labels a reader can follow. One drafted sentence often ends up part kept, part narrowed and part struck. Letting the reader see that happen is among the most useful things this review produces.

Keep the draft's exact words for anything you are about to change. Your paraphrase belongs nowhere near the record.

## When the material carries it outright

A statement is kept when the material carries it outright: a passage says it, and the boundaries the passage puts around it — who was measured, when, where, by whom, how firmly — are the boundaries the sentence implies.

Keeping is rarer than it looks. Most statements that feel obviously supported need narrowing once the passage is read closely, and a reviewer who keeps generously produces a document that fails precisely where somebody leans on it.

A kept statement carries its pointer in the sentence itself, beside the assertion rather than gathered at the foot of the page. Anyone checking one line should see which passage answers for it without reconstructing your reasoning.

## When the material carries less than the sentence claims

Narrowing is for statements the material nearly carries: the passage is on the subject but does not reach as far as the sentence, so the sentence is brought back until it does. Almost every case is one of five kinds of overreach, worth knowing by name.

There is the question of **who and where** — a passage that measured one segment, one territory, one firm's book, one class of arrangement has made that limit part of its finding, and the rewritten sentence carries the limit with it.

There is **when**. A figure belongs to the period it was measured in; another year's number is a different number, and a projection is not an observation.

There is **how firmly**. Sources hedge on purpose. *Suggests*, *may*, *early indications*, *is expected to* all mean something, and a draft that keeps the finding while dropping the hedge has quietly turned a possibility into a fact. The digits can be reproduced perfectly and the sentence still needs bringing down.

There is **who says so**. Material from a party with an interest in the answer — its own figures, its own study, its own announcement — remains usable and gets named in the sentence. Do not silently discard it, and do not let it read as independent work.

And there is **who did the arithmetic**. A quantity you calculated is not a quantity the material published. Carrying a base figure forward, adding two numbers, converting a proportion into a count: each produces something plausible that no passage states. Such a figure may appear only as a calculation, its inputs and method visible, never with a pointer as though a source had said it. When the draft's own numbers cannot all be true at once — a starting figure, an ending figure and a growth rate that do not reconcile — that failure is itself a finding, and it says something about where the figure came from rather than being a slip of the pen.

## Before writing the deliverables

Keep the claim split private: do not print a register, worksheet, analysis, or process narration. If the request itself flags a mixed-support compound sentence, audit that sentence first. In that case the log must visibly reconstruct the split: place every fragment, including the supported Tier-A fragment, in adjacent linked rows; classify a changed fragment as Tier B when the source supplies a faithful narrower or continuity statement, and Tier C only when no such statement survives. This exception overrides the log's normal omission of unchanged claims. Before drafting, test every available trio of earlier level, later level, and stated change. If they disagree, put the short equation and its decision consequence in the revised brief's opening paragraph even when an endpoint is unsourced. An excluded proposition may appear affirmatively only as its verbatim log entry; in the brief use negative evidence such as `No supplied source establishes ...`, never the original claim followed by a warning, bracket, or hedge. Use at most one compact sentence per draft subject. When the request says to classify every claim, the final log includes supported Tier-A propositions as well as changed ones; give fragments of a mixed sentence linked IDs such as `2a` and `2b`.

## Deciding between reducing a sentence and dropping it

This is where most wrong calls happen, and the question is never whether the number matches. It is whether the material measures the thing at all. A real figure reprinted without its year or its territory is narrowed, because the measurement exists and its conditions were stripped. A figure larger than anything any passage reports, or naming a quantity nobody measured for that period, is struck, because there is no smaller true version to retreat to. But where a passage does measure the same underlying subject and the draft has inflated it or shifted what is being counted, the statement is narrowed rather than struck, even though the draft's exact number appears nowhere in the set — brought back to the passage's own figure inside the passage's own boundary, with both measurements named so a reader can see what each one counts.

## Questions the packet must leave open

Some things the material leaves open, and the review's task is to pose them rather than close them.

The plainest case is two passages that disagree. Give both figures with their pointers and then say what the disagreement is made of — a different method, a different frame, a different period, a different definition of the thing being counted. Where a source's own account of its method explains the divergence, say so; where nothing explains it, say that instead. Never adopt the figure the draft happened to use, and never split the difference, because an averaged number is one that no passage contains and no reader can check.

A citation chain that stops also belongs here. A summary establishes what the summary reports, at the summary's own confidence; it does not stand in for a study nobody filed. Where the argument depends on methodology held in that absent study, say so and ask for it.

Thinness is separate from support, and escalates on its own terms. A finding resting on a few dozen respondents, one participant's records, or a single early cohort can be cited perfectly and still be far too narrow for the weight the draft puts on it. Keep it, name the base it rests on, and say what it will not carry.

And where the brief names the decision it feeds — a sum being committed, a vote being taken — a gap in the evidence stops being an audit note and becomes something that blocks the decision. Say it in those words. A figure without a named source, a stated period and a matching population is not fit to carry a recommendation, however confident the draft sounds.

## Claims with no footing in the packet

A removal can arise in three distinct ways; name the applicable one in the change record.

**The subject was not measured.** Search the complete packet, identify the nearest relevant passage, and state exactly what it covers and what the draft requires beyond it. An absence finding is useful only when it exposes the missing construct or boundary; a bare claim that support was not found gives the decision-maker nothing actionable.

**The record contradicts the draft.** Treat opposition as stronger than missing support. Quote the decisive source language, classify the proposition as contradicted, and remove it. Do not turn a reversed finding into a milder version of the draft merely to retain the topic.

**The proposition entered from outside the packet.** Familiar context can sound evidentiary even when it came from the reviewer: claims that a market is growing, regulation is tightening, or a practice is normal are common examples. However plausible, such material leaves the brief unless a supplied passage establishes it. The audit tests this record, not the reviewer's command of the field.

## Build the two deliverables

If most major draft subjects lack corresponding evidence, begin with a two-to-four-sentence source-coverage notice that counts and names the uncovered subjects, distinguishes the auditable subset, and then proceeds with the partial audit. For this incomplete-record case, make the accounting auditable by including the principal directly supported observation as Tier A in the log alongside every narrowed and unsupported proposition; do not confuse two sources covering different measures with a contradiction. Do not print a feasibility checklist or refuse the work. Otherwise the first characters returned are `## REVISED BRIEF: <draft title>`. There is no greeting, separator, method note, or restatement before either permitted opening. Always finish both the brief and the record; an unresolved conflict or a nearly empty brief is still a completed audit.

**Revised brief.** Use the draft's subjects, not a standard outline. Put the finding that most affects the named decision first. Each compact paragraph gives the strongest source-bounded statement with its pointer, carries the condition that limits it, and says a material absence as negative evidence: `No supplied source establishes ...`. Never leave the unsupported proposition standing as factual prose with a bracketed warning. Add at most two Open Questions, only when the answer could change the decision, and place them before the record.

Then write `---` and `## UNSUPPORTED-CLAIMS LOG`. Use the columns `# | Original claim (verbatim) | Tier | Specific evidence difference | Action`. Unless the assignment defines other labels, Tier A means carried at the same scope, Tier B means faithfully narrowed to a source ancestor, Tier C means removed because the proposition lacks an ancestor or is contradicted, and Hold open means a genuine source conflict. Log every B, C, and Hold-open proposition. If the task asks to classify every claim, also log A; whenever one drafted sentence splits across dispositions, use linked labels such as `3a` and `3b` and include the supported fragment so the split is visible.

Each reason names the pointer, what it establishes, and the exact changed field or inference. When compressing for length, compress sentences and rows, never the claim inventory.

## A real number can answer the wrong question

A particular failure hides inside quantitative claims and is missed more often than any other. The number is real, the source is real, and the source measures something *next to* what the draft claims — administrative cases where the draft says clinical, announced contracts where it says market, publicly reported figures where it says total, respondents to a survey where it says an industry.

Nothing about the sentence looks wrong, because the digits check out. The test is not whether the figure appears but whether the thing it counts is the thing the sentence asserts. Where they differ, the claim comes back to what the source actually counted, and the revised sentence names both — what was measured and what the draft made of it — so a reader can see the distance.

Two neighbours belong with it. Where a source expressly says it does not hold data on some question, a draft answering that question is unsupported and the disclaimer is the evidence, quoted. And where a source's own message runs opposite to the draft's framing — the draft calls a matter settled, the source describes it as open — that is not overstatement to be trimmed but a claim the record contradicts, corrected to what the source says with its words shown.

## Limits on neighboring evidence

Neighboring evidence has two limits; ignoring either one turns a justified removal into a false repair.

**A gloss is not a measurement.** Where a draft lays an editorial judgement over a sourced figure — calling a firm the clear leader, a level unprecedented, a position dominant — the figure may well be supported while the judgement is not. Do not narrow the sentence and keep the gloss in softened form. The number comes back at its own scope and the ranking or superlative leaves the brief entirely, because no passage ranked anybody. Self-reported provenance works the same way: a company's own figure about its own performance stays, named as such, but the characterisation the draft wraps around it does not travel with it.

**Plausible context is not adjacent evidence.** Where the draft asserts something the packet simply never addresses — the shape of a market, the direction of regulation, what has become standard practice — there is no nearby measurement to bring it back to, and treating it as scope drift invents a parent it does not have. Look for the passage; if none exists, say which source came closest and what it actually covers, and remove the claim. The adjacency rule applies only when a passage genuinely measures the same underlying thing. Where nothing does, the claim is unsupported, and calling it overstated credits it with evidence the packet never held.

## Labels the digits cannot prove

Statistical and administrative labels are claims, not decoration. `Not significantly different` does not by itself establish equivalence or practical comparability: without the primary design, an equivalence margin, effect size, and uncertainty interval, repeat the source's exact formulation and identify the unavailable evidence. Likewise, listed, provisionally accepted, permanently approved, announced, expected, and operational describe different states. A count mixing states cannot become a count of final outcomes; report the composition or say that the filed material does not separate it.

Cross-examine a draft against itself as well as against the sources. Whenever it supplies an earlier level, a later level, and a claimed one-interval change, calculate `(later / earlier) - 1`, even when the later level has no source ancestor. For multiple known intervals use `(later / earlier)^(1/n) - 1`. A material mismatch is the auditor's calculation, shown with a short equation in both the opening decision-relevant passage and the corresponding log row. When money, a vote, or another named commitment depends on the figure, say before the log that the mismatch blocks that premise until resolved.

## Check the finished audit

Trace every atomic proposition to one visible destination and rescan the final draft sentence. Confirm that unsupported or contradicted content appears only as verbatim log language or as a negative-evidence sentence, never as an affirmative claim beside a disclaimer. Check every quantity, status word, source pointer, qualifier, calculation, and linked split; keep conflicts unresolved and outside knowledge outside. Move any decision warning or question above the separator, then end on the complete final log row.