November 16, 2026

How to Check Whether Every Conclusion Matches the Results

A conclusion that outruns the results is the most expensive mistake an evidence team can ship. It survives review because the gap is rarely in the headline—it hides in a secondary claim, a confidence label, or a chart ca

Written byWisPaper TeamAI Research Workflow Team
Editorial illustration for How to Check Whether Every Conclusion Matches the Results

A conclusion that outruns the results is the most expensive mistake an evidence team can ship. It survives review because the gap is rarely in the headline—it hides in a secondary claim, a confidence label, or a chart caption that implies more than the underlying data showed. Checking conclusions against results is therefore not a final proofread. It is a structured pass through a document with one question repeated at every level: does this sentence say only what the evidence supports, and does the evidence we cite actually exist in the form we cite it?

The process below works for a client deck, a decision memo, a team handoff, or a project closeout. It assumes you already have a draft with sources, results, and conclusions. The job is to make the trace from result to conclusion explicit, flag every place where the link breaks, and know when the document is ready to submit.

WisPaper fits into this workflow at specific points: keeping the source base inside My Library so every claim can be re-opened against its original paper, checking direct questions against uploaded evidence with Library QA, and screening BibTeX entries through TrueCite when a reference list is part of the deliverable. The article uses those tools as part of the work, not as decoration.

What does “matches the results” actually mean?

A conclusion matches the results when three things are true at once. First, the direction is correct: the result says A increased, so the conclusion cannot quietly become “A improved outcomes” if the study measured only an intermediate metric. Second, the scope is correct: a finding from one population, one time window, or one dataset cannot be stated as a general fact about the whole market or field. Third, the confidence is correct: the author’s label—strong evidence, suggestive, exploratory—must reflect the design and sample, not the author’s enthusiasm.

NIH describes scientific rigor as strict application of the scientific method, including transparent design, methods, analysis, interpretation, and reporting. That definition gives the checking team a concrete standard. Transparent interpretation means a reader can see why the authors believe what they believe. Transparent reporting means the evidence trail is visible, not implied. When either is missing, the conclusion and the results have drifted apart even if no individual sentence is false.

The FAIR principles add a second lens. Research objects should be findable, accessible, interoperable, and reusable, with persistent identifiers, rich metadata, provenance, and clear usage conditions. In a deliverable, this translates into source provenance: every result node should name where it came from, when it was collected, and under what conditions. A conclusion that cannot be traced back to a source is not verified; it is asserted.

Is this a writing problem or an evidence problem?

Diagnose before you fix. Read the draft once and mark every claim as one of four types:

  1. Directly supported—the claim restates a result with the same direction, scope, and confidence.
  2. Overstated—the claim stretches direction, generalizes scope, or upgrades confidence beyond the result.
  3. Orphaned—the claim has no source attached, or the source attached does not contain the stated result.
  4. Synthesized—the claim combines two or more results into a new statement that no single source supports.

Most conclusion mismatches are type 2 and type 4. They feel true at a glance, which is why they pass internal review. The fix for type 2 is narrowing language. The fix for type 4 is either finding a source that states the synthesis directly, or recasting the synthesis as an explicitly labeled interpretation with its own confidence level.

If a draft has more than a handful of orphaned claims, stop editing and rebuild the source base. You cannot verify a conclusion against results that are not linked to their sources. Uploading the relevant papers and reports into My Library gives every later check a stable object to return to.

How do you set up a source-to-conclusion map?

A source-to-conclusion map is a table with one row per substantive claim. The columns are: claim ID, claim text, source, result as stated in the source, direction, scope, confidence label, and a pass/fail decision. Building this table is the single highest-leverage step in the whole workflow, because it forces the checker to write the result and the conclusion side by side.

The discipline is to quote or paraphrase the result from the source before looking at the conclusion. If you start from the conclusion, you will unconsciously search the source for anything that vaguely supports it. If you start from the result, you will see the gap.

Here is the rule that catches most mismatches: the conclusion may use different words than the result, but it may not change direction, broaden the population, extend the time frame, convert correlation language into causal language, or raise the confidence label. When any of those changes happens, the row fails.

For documents with many sources, use Research Projects to hold the working materials for this check in one place. That keeps the source map, the marked-up draft, and the final checklist together instead of scattered across chat threads and folders.

How do you verify confidence levels, not just content?

Content mismatches are visible: the result says 12%, the conclusion says market-leading. Confidence mismatches are subtler because they are often conveyed by adjectives rather than numbers. “The data show,” “the evidence suggests,” and “initial findings indicate” are three different confidence levels wearing the same sentence structure.

The check is to assign every conclusion one of three explicit levels and ask whether the underlying design justifies it:

  • Established: multiple independent sources converge; the result has been reproduced or triangulated; scope is broad enough for the claim.
  • Supported: a single strong source directly states the result; scope matches the claim.
  • Indicative: the result comes from a pilot, a small sample, a single case, or a preliminary analysis; the claim must carry that limitation.

A common failure is attaching “Supported” confidence to a result that is really “Indicative” because the author wants the recommendation to feel stronger than the evidence allows. The fix is not to soften every claim into mush; it is to be honest about the level and let the recommendation logic account for it. A decision can still be made on indicative evidence; the memo just has to say that is what it is.

When a confidence question turns on whether a cited reference actually says what the draft claims, run the BibTeX through TrueCite to check the references against real academic databases. A reference that does not resolve is a confidence problem even before you read the paper.

What does a worked example look like?

Here is a decision memo example reduced to the checking exercise. The deliverable is an internal memo recommending a pricing change for a B2B analytics product, based on three internal analyses.

Result 1: Among 42 pilot accounts, 31 increased logins within 30 days after the pricing change. The pilot was self-selected; accounts volunteered to join.

Draft conclusion: “The pricing change increased product engagement across the customer base.”

Diagnosis: Overstated in both scope and causality. The pilot included 42 self-selected accounts, not the full base, and the design cannot separate causation from voluntary participation. Direction is positive, but the claim broadens population and implies causality.

Revision: “Among 42 self-selected pilot accounts, 31 logged in more often within 30 days of the pricing change.” Label: Indicative. This revision matches result to conclusion, keeps the positive signal, and removes the unsupported generalization.

Result 2: Support tickets mentioning price confusion fell from 18 per week to 9 per week in the pilot month, the first measured change in a year.

Draft conclusion: “Customers understand the new pricing structure better.”

Diagnosis: Synthesized and slightly overreaching. Ticket volume is a proxy for confusion, not a direct measure of understanding. The decline could reflect fewer price-related questions for other reasons. The direction is supported; the interpretation of the cause is not direct.

Revision: “Price-confusion support tickets fell by half in the pilot month; this is the first change in that metric in a year and may reflect reduced pricing friction.” Label: Supported for the ticket decline, Indicative for the interpretation. This is the kind of split confidence that a good source-to-conclusion map surfaces: one sentence, two claims, two different levels.

The memo passes when every substantive row in the map shows direction match, scope match, and an honest confidence label. If one row fails, the memo does not ship until that row is revised or the conclusion reclassified as an explicitly labeled interpretation.

What checkpoints should the team hit?

Checkpoints turn the workflow from an individual exercise into a reproducible process. Run them in order.

  1. Source audit: every claim of substance has a source attached, and the source is actually present in the library or file set.
  2. Result extraction: results are copied or paraphrased from sources before conclusions are read.
  3. Direction and scope pass: each conclusion is compared against its result for direction, population, time frame, and causal framing.
  4. Confidence pass: each conclusion receives an explicit level, and any level above Indicative is justified by design.
  5. Reference screen: all citations resolve to real sources in the expected form, including author, year, venue, and identifier.
  6. Traceability pass: a fresh reader can find the source for every conclusion in under a minute.
  7. Readiness decision: the deck, memo, or closeout is marked ready, revise, or rebuild based on the pass/fail counts.

A document can ship with a small number of Indicative claims if they are labeled and not load-bearing. It cannot ship with any orphaned or overstated claims. That is the completion criterion, stated plainly.

How does this fit a submission or closeout schedule?

Schedule the conclusion check as its own activity, not as a tail on the writing day. For a client deck due Friday, run the source-to-conclusion map on Wednesday afternoon, the confidence and reference passes Thursday morning, and leave Thursday afternoon for revisions and the readiness decision. For a project closeout, set the same sequence against the closeout review date.

ICMJE tells authors preparing a biomedical manuscript to follow the selected journal’s current instructions for article type, format, required statements, and submission materials. The conclusion check is the layer underneath those requirements: before you format the required statements, every conclusion in the manuscript has to match the results as reported. Running the check before formatting saves a full revision cycle.

For handoffs between teams, the source-to-conclusion map itself is a valuable deliverable. The next team inherits not just the conclusions but the evidence trail, which is exactly what makes the handoff traceable rather than verbal.

How do you know it is ready?

The document is ready when four conditions hold simultaneously. Every substantive conclusion traces to an identifiable source in under a minute. Every conclusion preserves the direction and scope of its source result. Every conclusion carries an explicit confidence level that the design supports. And no cited reference exists only in the author’s memory of the literature.

If a team can answer “show me where this came from” for any sentence in the deck without opening a search browser, the traceability condition is met. If the answer starts with “I remember reading something,” it is not.

The final state is not a document with more caveats. It is a document where the reader trusts the strong claims because the weak claims are honestly labeled. That is the difference between a draft and a submission.

FAQs

For a 15–20 slide deck with a single primary source set, the full source-to-conclusion map takes two to four focused hours, split across the seven checkpoints. The bulk of the time goes into result extraction and the confidence pass; revision time is the variable. Teams that keep sources organized in one place, such as My Library, spend less time hunting for the original paper behind each claim.