Capstone: the research memo

Read this as an inventory check. Anything missing is tonight's homework, not a reason to skip the memo:

• b2 - the decision-scoped claim list. Claims that could be false, tied to a decision with a date. This becomes the memo's spine.
• b3 - the documented search and tier tags. Where you looked, with what queries, and what tier every source carries. This becomes the first half of the method appendix.
• b4 - the screening log with exclusion reasons. What you excluded and why. The reasons are the valuable part, because they are what a reviewer cannot otherwise reconstruct.
• b5 - the extraction table. One row per source with sample, measure, number, location, funder and population. Every figure in the memo traces to a row here.
• b6 - the audit verdicts and the triage. Which claims survive, which failed, and whether each failure was fatal, work or honesty.
• b7 - the contradiction and replication record. Who disagrees, how strong the disagreement is, and whether the finding has held up more than once.
• b8 - the calibrated claim statements. Claim, weight, basis, and what would change your mind. These are the sentences that actually go in the memo.
• b9 - the qualitative audit trail. Codebook version, coder agreement, saturation evidence, verbatim quotes, consent and data handling.

Notice what is not on the list: a literature summary. Nobody is going to read a summary of the literature. They are going to read your claims and then check one of them.

A memo ships when it passes two gates. Everything else in this session is machinery for getting through them.

• Gate one: every load-bearing claim either survives the audit or is explicitly marked unknown. There is no third state. A claim that failed and got softened into "early evidence suggests" has not passed the gate - it has been smuggled through it wearing research clothes. Load-bearing means the decision changes if the claim is false; anything that is not load-bearing can simply be cut.
• Gate two: the memo names what would change its mind. Per claim, in the confidence sentence, and once at the top for the recommendation as a whole. This is the difference between a position and a preference, and it is the line experienced readers look for first.

Both gates are cheap if you built the artifacts and expensive if you did not, which is the entire argument for building them as you go.

The unknown section is the one most likely to be edited out by someone senior who wants a cleaner document. Write it so that removing it obviously costs something:

Three ingredients make it survive: the price of not knowing, a test that is small enough to actually be commissioned, and a decision rule for the interim. An unknown with a cost, a test and a rule is a plan. An unknown without them is a caveat, and caveats get deleted.

• Re-run the seven audit rules on the final claim set. Claims move during writing. A claim you audited as "institutions most likely hold the budget" has a habit of becoming "institutions hold the budget" by the third draft, and that edit changes the verdict. Audit what you are actually sending.
• Re-check every number against its recorded location. Not against the draft, not against your memory of the source, and not by asking a model to confirm it. Open the extraction row, open the source at the recorded location, read the figure. This is the check that catches the denominator that quietly disappeared.
• Confirm no citation has been retracted. Crossref carries retraction and post-publication update information, and it needs no sign-up. Run the final reference list through it. A withdrawn citation is a credibility loss you cannot argue your way out of, and it is the cheapest possible thing to prevent.

• Quotes and confidentiality. Every participant quote verbatim from the transcript, with its reference. And a hard look at what went where: which tools saw the transcripts, which saw confidential client material, and whether any of that breached what the participants agreed to or what your contracts allow.
• The disclosure matches the venue's rule, and the rules genuinely differ. ICMJE asks for it in both the cover letter and the submitted work, splitting writing assistance into acknowledgements and data or analysis work into methods. COPE puts it in Materials and Methods. Springer Nature documents it in Methods. Science asks for the cover letter plus methods or acknowledgements. Elsevier requires a separate declaration statement with prescribed wording. For commercial insights work, ICC/ESOMAR Article 9(b) requires disclosing whether AI or synthetic data played a significant role in sampling, deployment, analysis or interpretation, and to what extent human oversight was involved. Write to your venue, not to a generic template.
• Read the confidence words out loud. Out loud, genuinely. "Institutions hold the budget" and "institutions most likely hold the budget, based on national IT-spend statistics and twelve buyer interviews" are different sentences, and your ear catches the mismatch between the word and the evidence faster than your eye does.

Four moves, then stop talking:

• The decision. "You are deciding by the 30th whether we build lecture notes for higher education." One sentence, and it makes everyone in the room the same kind of listener.
• The recommendation with its confidence. "We recommend a limited pilot rather than a full build. Moderate confidence." The confidence word is not hedging, it is calibration, and stating it up front buys you enormous credit later.
• The two or three claims that carry it. Not all nine. The ones the recommendation would break without, each with its basis in half a sentence.
• The unknown and its test. "One thing we do not know, here is what it costs us, here is the two-week test." Ending on this rather than on your evidence base is what makes the room trust the rest.

The method appendix goes behind the recommendation, not in front of it. It exists so that when someone asks how you know, the answer is a page reference rather than a story. Bring it, refer to it, do not present it.

They will, and they will find the right one, because weak claims read as weak even to people who have not done the work. There is exactly one good response: agree with them, immediately, and point at the audit row.

"You are right, and that is the one I am least confident in. It rests on a single study of 48, it is in the memo marked as untested, and the two-week benchmark on page four is how we would settle it."

Watch what that does. The person who came to find a hole discovers you found it first, documented it, and priced it. Everything else in your memo just became more believable, because you have demonstrated that you say so when something is thin. A pre-empted weakness is a credibility gain. A defended weakness is a credibility loss, and it costs you the strong claims too.

The version to avoid, and it is tempting under pressure: adding evidence in real time from memory. Any number you produce in a meeting without its recorded location is a number you may have to retract in writing afterwards.

• "Where did that number come from?" Source, sample, and the location in the source. If the answer takes longer than ten seconds, your extraction table is not doing its job.
• "Did you look at anything that disagreed?" Name the disagreeing source and say how you weighted it. This is the b7 record, and having it is the difference between synthesis and advocacy.
• "What would change your mind?" You wrote it into every confidence sentence, so read it back. The room is not testing your evidence at this point, it is testing whether you have a position or a preference.