From rows to a narrative 7 min live
You have 22 sources, an extraction table and nine surviving claims. The temptation is to write the table up in source order, one paragraph per paper, and call it a literature review. That produces a document nobody can act on and nobody can audit, because the reader has to do the synthesis themselves. The order that works instead is decision-first, and it is four moves.
LiveThe order that works, in four moves5 min▶
- Group claims by the decision they serve, not by source. Cadence's decision has four parts: who holds the budget, what consent recording requires, how long procurement takes, and whether the engine transfers. Every claim goes under one of those headings. If a claim does not fit any of them, it is interesting and it does not belong in this memo.
- State each claim with its weight. Not "several studies suggest" but a stated certainty level and the basis for it. A reader should be able to tell, from the sentence alone, whether to plan on it or hedge against it.
- Keep disagreement visible instead of averaging it away. "The evidence is mixed" is the phrase that destroys the most value in research writing. It sounds balanced and it deletes everything you learned in b7: the size of the split and which side the stronger studies fall on.
- Put the unknowns in the body, not a footnote. If a decision rests partly on something you could not establish, that belongs in the paragraph where the decision is discussed. Unknowns moved to the back matter are unknowns nobody reads before committing budget.
Self-studyWhy a synthesis is not a summary3 min read▶
A summary compresses what sources said. A synthesis states what you now believe and why. The difference is not stylistic, and it is not about length. Only the second one can be audited, because only the second one makes a claim that could be wrong.
"Smith (2023) found institutional purchasing dominant in higher education software. A 2024 student survey reported strong individual willingness to pay. A sector procurement study described cycles of six to twelve months. Views on transferability of speech models across domains vary."
"We should plan for institutional procurement with a six to twelve month cycle. Three independent procurement sources support this and two smaller vendor-run student surveys point the other way; we weight the procurement sources higher on independence and sample. The one thing we cannot yet establish is whether our engine transfers to lecture audio, and that is the item we are pricing separately below."
Every sentence in the second version is falsifiable, which is what makes it defensible. Ask of any paragraph you write: could a reviewer disagree with this, and would they know where to aim? If the answer is no, you have written a summary.
Confidence language in four parts 8 min live
A defensible claim has four parts, and most writing ships with one or two of them. Claim: what you believe. Weight: how strongly, in vocabulary that means the same thing to you and your reader. Basis: what the weight rests on. Revision trigger: what would change your mind. The fourth part is the one people leave out, and it is the one that converts a memo from an argument into a working instrument.
LiveBorrowing GRADE's vocabulary so "moderate" means something5 min▶
The problem with confidence words is that they are private. Your "likely" and your reader's "likely" are different numbers, and neither of you finds out until the decision goes wrong. GRADE solves this for evidence certainty with four levels and published wording, and you can borrow it directly even when you are writing a market memo rather than a clinical review:
- High - "We are very confident that the true effect lies close to that of the estimate."
- Moderate - confident, with reservations you can name.
- Low - "Our confidence in the effect estimate is limited."
- Very low - little confidence; the estimate is a placeholder.
What makes the vocabulary usable is that GRADE also names the reasons a body of evidence gets downgraded, so your weight is arguable rather than asserted. The five downgrading factors are risk of bias, inconsistency, indirectness, imprecision and publication bias. In a memo that becomes one clause: "moderate rather than high, because the studies are indirect - two of the four are on corporate procurement rather than higher education." A reader can now argue with your weight, which is exactly what you want. An unarguable confidence level is not a confidence level; it is a mood.
Self-studyThree versions of the same claim4 min read▶
Same evidence, three sentences. Only the third one survives a reviewer, and only the third one can be updated when new evidence arrives.
"Institutions, not students, hold the budget for lecture-notes tools, so we should sell to institutions."
"Evidence on budget ownership in higher education is mixed, with some sources suggesting institutional purchasing and others pointing to student willingness to pay. Further research would be beneficial."
"We believe institutions rather than students hold the budget for lecture-notes tools (claim). We hold this with moderate certainty, not high, because two of our four supporting sources study corporate rather than academic procurement (weight, with the downgrade named). It rests on three tier-1 procurement sources and one sector body, S04, S09, S11 and S17; two smaller vendor-run student surveys disagree and we weight them lower on independence and sample (basis, with the disagreement visible). Three institutional buyers telling us students pay directly would move us off this position (revision trigger)."
The middle version is the dangerous one, because it feels like rigour. It has a weight with no basis, a disagreement with no weighting, and no revision trigger, so nobody can act on it and nobody can challenge it. "Further research would be beneficial" is the tell.
Writing the unknown, and saying what you used 8 min live
Nine claims survive. One does not, and it never will, because a single vendor-funded preprint with n=48 is not evidence that our engine transfers to lecture audio. That tenth claim is not a hole in the memo. Written properly it is the most credible paragraph in it, and it follows a fixed four-part pattern. Then the part that gets researchers in actual trouble: where you disclose the AI you used, which every body requires and no two bodies agree on.
LiveThe honest-unknown pattern4 min▶
- Mark it unknown. Plainly, in the body, in the section where the decision is discussed. "We could not establish whether our engine transfers to lecture audio."
- Price what not knowing costs. If we are wrong, what does it cost - retraining time, a delayed launch, a refunded pilot? An unknown with no price attached gets treated as a rounding error.
- Name the test that would settle it, and its cost in days. "Ten lecture recordings across three subjects, transcribed and scored against our existing benchmark: four days of engineering time." This is the sentence that turns an unknown into a plan.
- Say what you will not do until then. The commitment boundary. "We will not put lecture-notes accuracy in marketing material, and we will not quote an accuracy figure to a pilot institution, until that test has run."
Executives do not punish unknowns. They punish surprises. A marked, priced unknown with a named test reads as command of the problem. The same gap discovered later, in a paragraph that had implied certainty, reads as something you missed or hid. The four-part pattern costs you a paragraph and buys you the right to be believed on the other nine claims.
LiveDisclosure - everyone requires it, nobody agrees where4 min▶
Two things are settled. AI cannot be an author anywhere - it cannot take responsibility for the accuracy, integrity and originality of the work, and as a non-legal entity it cannot declare conflicts of interest. And every major body requires you to disclose use. What they disagree on is the location and the wording, which means "I disclosed it" is not a complete answer. You disclose it where your venue says.
One venue goes further than the rest: Science states that sources cited in its journal content "may not be authored or coauthored by AI tools". That is a constraint on your bibliography, not just your methods section, and it is worth checking before you cite an AI-generated report.
| Venue | Where the disclosure goes | What it must say |
|---|---|---|
| ICMJE | Both the cover letter and the submitted work | Writing assistance in acknowledgements; data collection, analysis or figures in methods. Nondisclosure "may require corrective action and may be construed as misconduct" |
| COPE | Materials and Methods | How AI was used. AI tools "cannot take responsibility for the submitted work" and cannot assert the presence or absence of conflicts of interest |
| Springer Nature | Methods | Documented use. LLMs "do not currently satisfy our authorship criteria"; generative-AI images largely not permitted, narrow exceptions, labelled |
| Science / AAAS | Cover letter plus methods or acknowledgements | Full description of use. AI may not be an author, and cited sources "may not be authored or coauthored by AI tools" |
| Elsevier | A separate published declaration statement | Prescribed wording. Basic grammar and spell checking needs no declaration; primary research images from AI are not permitted |
| ICC/ESOMAR Code 2025 commercial research | Art 9(b) to the record; Art 7(e) to the client | Whether AI or synthetic data "played a significant role in sampling, deployment, analysis, or interpretation of the data, and to what extent human oversight was involved" |
Self-studyWriting the disclosure sentence for commercial work3 min read▶
If your output is a client deliverable rather than a journal article, the ICC/ESOMAR Code 2025 is the instrument that binds you, and it asks for two things: disclosure of significant AI involvement across sampling, deployment, analysis and interpretation, and a statement of the extent of human oversight. Article 7(e) makes the client-facing half explicit. Write it once, keep it as a template, and fill in what actually happened.
"AI tools were used in this study as follows. Literature search and first-pass screening: AI-assisted across 62 candidate sources, with all 22 included sources read and verified by the named researcher. Extraction: AI-assisted into a structured table, with every figure checked against the source page. Analysis and interpretation: human only. Synthetic respondents or synthetic data: none. Human oversight: the named researcher reviewed and is accountable for all included sources, every extracted number, and all claims and confidence levels in this report."
Notice that the useful disclosure is stage by stage, because "AI was used in preparing this report" tells a client nothing about which stage to trust. The oversight sentence at the end is the one the Code is really after, and it is also the one that makes you check whether the statement is true before you sign it.
Draft the memo body, live ★ 16 min · everyone writes
The lab below is at the top of the ladder: all five levers on, 9 of 10 claims surviving, 22 sources. Nothing left to toggle. This session's work is writing, and the lab is your source of truth for it - each survivor row shows the tier and the source ids that justify the weight you are about to claim. One calibrated sentence per survivor, then the honest-unknown paragraph for the tenth.
Read the nine survivor rows and group them under the four parts of the Cadence decision: budget ownership, consent for recording, procurement timing, engine transfer.
For each survivor, read its tier and source ids off the row. That is your basis clause, and it is the reason this session is assembly rather than invention.
Assign a weight using the GRADE vocabulary, and name the downgrade where you use one - indirectness and imprecision will do most of the work here.
Write the four-part sentence: claim, weight with its reason, basis with the disagreement visible, revision trigger. Nine sentences. Where a claim was qualified in b7, keep the split in the sentence.
Now the tenth. Write the honest-unknown paragraph for engine transfer: mark it, price it, name the test and its days, and state what you will not claim until it runs. Read it back and check that it does not read as an apology.
Try it yourself - this week ◐ 35-45 min total
- Take a memo or report you have already sent and mark up its load-bearing paragraphs for the four parts. Count how many have a revision trigger. For most people the answer is zero, and that is the habit this session is trying to change.
- Find one "the evidence is mixed" sentence in your own writing and rewrite it with the split visible: how many sources on each side, how strong they were, and which way you lean because of it.
- Write the honest-unknown paragraph for a real gap in your current work, all four parts, including the cost in days of the test that would settle it. Put it in the body of the document, not the appendix.
- Look up the disclosure rule for the venue or client you actually publish to, and draft your stage-by-stage statement now, while you still remember which stages used AI. Keep it as a template.
Official sources covered
This session borrows its confidence vocabulary from an evidence-grading standard and its disclosure rules from the publishing and market-research bodies. Every quoted phrase below comes from the body named. This page covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What is the difference between a summary and a synthesis?
The distinction is about who does the thinking. A summary hands the reader compressed source material and leaves the judgement to them. A synthesis makes a claim that could be wrong, with its weight and basis attached - which is exactly why a reviewer can test it and a decision-maker can act on it.
2 · Claim, weight, basis and one more. What is the fourth part of a calibrated confidence sentence?
The revision trigger is the part almost everyone omits, and it is what turns a memo into a working instrument: when the named evidence appears, the reader knows the position has moved without having to re-litigate the whole document. "Further research would be beneficial" is the mush version of the same instinct and commits to nothing.
3 · You used AI to help screen and extract. Why is "I disclosed it" not a complete answer?
They agree AI cannot be an author and that use must be disclosed, and they diverge on location and wording. Science additionally requires that cited sources not be authored or coauthored by AI tools, and for commercial work the ICC/ESOMAR Code wants the significant-role disclosure plus a statement of the extent of human oversight. Check your venue before you write the sentence.