Scroll anchoring and CSS anchor positioning share a word but solve different problems. Scroll anchoring is a browser mechanism that adjusts a scroll offset to keep the user’s view stable when layout changes above it. Anchor positioning lets an author position one box relative to another. One changes the scroll position; the other determines box geometry.
This guide develops one focused reader question within the Websites subject guide. It adds a complete mechanism, original worked decisions, changed-context transfer checks and explicit limits. Read it with The Gold Standard of Education | Learning That Transfers into Life and the How X Works.
All numerical examples are invented teaching cases. They are not operational measurements, production records, financial advice, conservation findings or compliance results.
A fictional reader is viewing a paragraph whose top is 300 CSS pixels below the scrollport top. A late image inserted above it adds 120 pixels of layout height. Without scroll anchoring, the paragraph moves to 420 pixels. With a 120-pixel scroll-offset adjustment, it remains visually at 300. Separately, a tooltip positioned 12 pixels above a button uses CSS anchor positioning; that relationship does not select the scroll anchor node. The CSS Scroll Anchoring specification defines anchor-node selection and overflow-anchor. The CSS Anchor Positioning Working Draft defines position-anchor, position-area and related geometry.
Continue with Tracing Anchor-Position Fallbacks When the Preferred Placement Overflows, Finding the Containing Block for an Absolutely Positioned Element and the Websites root.
The short answer is to keep the target quantity, observation and reporting rule separate. Name what the system intends to estimate, what was actually observed, how the summary was formed and which conditions make the inference defensible. If one field is missing, keep the ambiguity visible.
eduKateSG · Websites
Find your next learning step
Scroll anchoring and CSS anchor positioning share a word but solve different problems. Scroll anchoring is a browser mechanism that adjusts a scroll offset to keep the user’s view stable when layout changes above it. Anchor positioning lets an author position one box relative to another. One changes the scroll position; the other determines box geometry.
Chapter contents
Frame the question · Chapters 1–5
2. Separate the two target quantities
Work the example · Chapters 6–10
6. Work the central calculation
7. Check units before interpretation
Check the evidence · Chapters 11–14
11. Use controls that can fail differently
12. Carry uncertainty into the claim
Transfer the method · Chapters 15–18
16. Change one consequential condition
Audit and continue · Chapters 19–22
19. State what it cannot establish
20. Repair an overconfident sentence
Ask whether the symptom is a changed scroll offset, a shifted content box or a positioned overlay moving relative to its anchor.
Write an evidence ledger before calculating: target quantity, object, unit, boundary, observation rule, transformation and decision limit. Mark every input as observed, stipulated, fitted or assumed.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “define the reader question”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Scroll anchoring selects an anchor node for a scrolling box; anchor positioning resolves an absolutely positioned box against anchor references.
Draw the mechanism in order. Separate the underlying state, the measurement, the mathematical summary and the claim. Then name one rival explanation and the additional observation that could distinguish it.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “separate the two target quantities”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Name scroll container, anchor node, overflow-anchor state, layout insertion, position-anchor and containing block.
Keep the comparison fair by holding the specimen, interval, protocol and denominator constant. If one changes, state that the new number answers a different question even when the label looks familiar.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “name the observation rule”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Record before-and-after rectangles, scrollTop, inserted height, browser version, writing mode and any scroll snap.
Use two checks that can fail differently: reconstruct a total, test a limiting case, repeat the result from definitions or inspect a shared control. Two calculations that reuse one assumption are only one check.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “protect the denominator and boundary”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Build the 120-pixel insertion case and a separate tooltip twelve pixels above its button.
Separate measurement from decision. A correct number can still be too local, uncertain or method-dependent for the intended action. Report it with its conditions and say what stronger evidence would be needed.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “build the fictional evidence ledger”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
The browser can add 120 to scrollTop to keep the paragraph stable while the tooltip geometry remains a different calculation.
| Mechanism | Input | Output |
|---|---|---|
| scroll anchoring | layout shift above view | scroll-offset adjustment |
| anchor positioning | anchor and positioning rules | positioned-box geometry |
Preserve identity and order. Equal totals can hide different mechanisms when records come from different specimens, resources, homes or viewing geometries. Aggregate only after the question no longer depends on that identity.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “work the central calculation”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
CSS pixels and device pixels differ; use one coordinate space when reconstructing the scroll adjustment.
Run a changed-context transfer check. Change one consequential condition while holding the remaining basis fixed, calculate again and explain why the result moves or stays stable.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “check units before interpretation”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Scroll anchoring is commonly automatic, while overflow-anchor can exclude candidates. Anchor positioning needs author rules.
Put the boundary of inference beside the answer. Name the system, representation and interval to which the result applies, then name the nearest tempting conclusion the evidence does not establish.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “keep method settings visible”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
A usable diagnosis records the selected scroll container and the relevant positioned containing block separately.
Use uncertainty as structure. Vary the input most capable of changing the conclusion over a defensible range and show whether the ranking or interpretation changes.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “locate the usable range”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Scroll snap, programmatic scrolling and suppression triggers can limit or suppress scroll anchoring behavior.
Finish with a reproducible evidence package: source records, definitions, exclusions, calculations, validation result and revision rule.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “recognise saturation and censoring”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Check scrollTop, element rectangles, Layout Shift evidence and a reduced case with overflow-anchor:none.
Write an evidence ledger before calculating: target quantity, object, unit, boundary, observation rule, transformation and decision limit. Mark every input as observed, stipulated, fitted or assumed.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “use controls that can fail differently”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Carry asynchronous image decoding and font changes into the timeline; several shifts can accumulate.
Draw the mechanism in order. Separate the underlying state, the measurement, the mathematical summary and the claim. Then name one rival explanation and the additional observation that could distinguish it.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “carry uncertainty into the claim”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
A stable visual paragraph does not imply no layout change; the browser may have compensated with a scroll adjustment.
Keep the comparison fair by holding the specimen, interval, protocol and denominator constant. If one changes, state that the new number answers a different question even when the label looks familiar.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “separate mechanism from appearance”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Compare runs only with aligned viewport, initial scroll, loading order, content size and browser state.
Use two checks that can fail differently: reconstruct a total, test a limiting case, repeat the result from definitions or inspect a shared control. Two calculations that reuse one assumption are only one check.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “compare like with like”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Test the rival explanation that application code restored a saved scroll position instead of native scroll anchoring.
Separate measurement from decision. A correct number can still be too local, uncertain or method-dependent for the intended action. Report it with its conditions and say what stronger evidence would be needed.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “test a rival explanation”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Set overflow-anchor:none on the scrolling box in the fictional case. The paragraph now moves from 300 to 420 when the image expands.
Preserve identity and order. Equal totals can hide different mechanisms when records come from different specimens, resources, homes or viewing geometries. Aggregate only after the question no longer depends on that identity.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “change one consequential condition”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
The CSS-positioned tooltip can still remain twelve pixels from its button, proving the two mechanisms can vary independently.
Run a changed-context transfer check. Change one consequential condition while holding the remaining basis fixed, calculate again and explain why the result moves or stays stable.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “recalculate the changed case”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
The example establishes separate state changes for scroll stabilization and positioned geometry.
Put the boundary of inference beside the answer. Name the system, representation and interval to which the result applies, then name the nearest tempting conclusion the evidence does not establish.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “state what the result establishes”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
It does not guarantee selection of a particular anchor node, freeze draft behavior or assess every browser.
Use uncertainty as structure. Vary the input most capable of changing the conclusion over a defensible range and show whether the ranking or interpretation changes.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “state what it cannot establish”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Replace “the CSS anchor stopped the page jump” with the measured scroll-offset adjustment and the independent positioned-box rule.
Finish with a reproducible evidence package: source records, definitions, exclusions, calculations, validation result and revision rule.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “repair an overconfident sentence”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Is overflow-anchor for tooltips? No. It controls scroll-anchoring participation.
Does position-anchor stabilize scroll? No.
Can both operate together? Yes, on different state.
Write an evidence ledger before calculating: target quantity, object, unit, boundary, observation rule, transformation and decision limit. Mark every input as observed, stipulated, fitted or assumed.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “answer focused reader questions”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.
Use both W3C specifications and preserve a time-ordered geometry trace before changing layout code.
Draw the mechanism in order. Separate the underlying state, the measurement, the mathematical summary and the claim. Then name one rival explanation and the additional observation that could distinguish it.
Apply the rule to the central fictional case with unrounded inputs, then label the result with its object, unit and protocol. Name one alternative explanation and the observation that would distinguish it.
For Websites, complete seven fields: target, object, unit, boundary, protocol, result and limit. For “continue through verified routes”, a plausible number can otherwise hide a changed definition or undocumented condition.
Change one condition that matters while holding the remaining basis fixed. Recalculate or re-inspect the evidence, state what stays comparable and name the stronger conclusion that still does not follow.
Close with two lines. The first states what the evidence establishes under the named conditions. The second states the nearest tempting conclusion that remains unsupported. Ask another reader to identify the condition that would make the first line false.
Evidence check: preserve the raw observation, the transformation, the comparison rule and the failed alternatives in one audit note. A future reader should be able to reproduce the narrow result without guessing which conditions were held constant.