The columns that earn their place 8 min live
One row per source, no exceptions, including the sources that turned out to disagree with you. A table with a row missing is not a table, it is a shortlist. Twelve columns cover almost every question anybody will ask you: source id, full citation, year, design or type, tier, population, sample size, the measure as the source itself defines it, the number, the location in the source, funder or interest, and a verbatim quote for anything you intend to state as a finding. Each one exists because it catches a specific failure, and once you have seen the failure you stop resenting the column.
LiveWhy the location column is the one that saves you4 min▶
Every column on that list is useful. One of them is load-bearing. Location in source - page, table number, figure number - is what makes a number re-checkable, and re-checkability is the entire difference between evidence and a rumour with a citation attached.
Watch what happens without it. You record "38%, Smith 2024". Six weeks later somebody asks whether that is weekly or monthly use. You open the PDF. It is 31 pages. The number appears twice, once in the abstract and once in a subgroup table, and they do not mean the same thing. You are now doing the extraction again under time pressure, in front of the person who asked - and the honest answer, "let me come back to you", costs you the room.
Now watch what happens over a longer horizon. A number without a location gets copied from your table into a slide, from the slide into someone else's deck, from that deck into a strategy doc. At each hop it loses a qualifier: the population narrows or widens, "weekly" becomes "regular", the confidence interval evaporates, 38% becomes "nearly 40%" becomes "around half". Nobody lied. Nobody could check. A number nobody can re-check will drift, and it drifts in the direction the reader wants.
The habit is small: when you write the number, write where you got it in the same keystroke. Page and table, every time, including for your own internal data.
Self-studyThe measure column, and the definition swap3 min read▶
The subtlest column is measure, as the source defines it. Not as you would define it. Not as your product team defines it. As it is written in the paper.
The Cadence question is full of these. You want to know about "adoption". One source measured "has ever opened the tool", another "used it in a typical week", another "used it for at least three lectures in a term". All three would be reported to you as adoption, and all three are legitimate. Put them in the same column without their definitions and you have built a comparison that means nothing while looking rigorous - which is worse than no comparison at all, because it survives a skim.
The rule: copy the definition, then decide whether the sources are comparable. Sometimes they are not, and the finding is "three studies measure adoption three different ways, so the range 12% to 38% is not a range". That sentence is more valuable than an average, and you can only write it if the definitions are sitting in your table.
Extracting with AI without laundering its errors 8 min live
Extraction is the second of the two steps AI genuinely compresses, and the compression is enormous - a job that took a day takes twenty minutes. The workflow that keeps it honest has three parts and the third is not optional. Give the tool your column definitions, so it fills your table rather than inventing its own. Extract into the table. Then verify every number that will appear in the memo against the source itself. Not a sample. Every number you will state. The verification is where the time saving gets spent, and it is still a massive net win.
LiveWhere AI extraction actually goes wrong5 min▶
These are not hallucinations. Every one of the failures below returns a number that is in the paper, which is precisely why they get through review:
- Wrong denominator - 38% of respondents becomes 38% of students, or 38% of institutions. The figure is right; the base has changed underneath it.
- A subgroup reported as the whole - the number from the postgraduate column arrives in your row as the study's headline finding.
- A baseline taken as an outcome - the pre-intervention measure extracted as the result. Common with anything reported in a before-and-after table.
- Units silently converted - months to weeks, per-user to per-seat, percentages to percentage points. No flag, no note, a plausible number.
- An abstract figure the full text qualifies - the abstract says the intervention improved recall; page 14 says it did so in one of three conditions and not at follow-up. The tool read the abstract.
Notice the pattern: every failure is a confident restatement of something true about a different thing. No verification pass that reads only the extracted table can catch any of them. You have to go back to the source, which is what the location column makes cheap.
"Pull out the key findings and statistics from these 24 papers on note-taking tools in higher education, and put them in a table I can use."
"Extract these papers into a table with exactly these columns, and leave a cell empty rather than filling it with an inference: source id · full citation · year · study design as the paper labels it · population, including country and institution type · sample size (state the denominator the reported figure uses) · outcome measure copied word for word from the paper · the number · location of that number (page, and table or figure) · funder and any declared competing interest · one verbatim quote, 25 words maximum, containing the number. Do not convert units. Do not round. If the number appears in the abstract and the full text qualifies it, use the full text and say so in the quote column."
"What proportion of students roughly use AI note-taking tools, according to these papers? An approximate figure is fine."
"For each paper that reports a proportion of students using note-taking tools, give me the figure exactly as the paper states it - same precision, same units, same denominator, no rounding and no averaging across papers. Include the measure definition and the page. If a paper does not report a proportion, write 'not reported' rather than estimating one."
Self-studyTwo tool limits that matter at this step3 min read▶
Elicit's Extract Data workflow does exactly this job and does it well: you define columns, it fills them across your paper set. The limitation from b4 applies with more force here than anywhere else in the pipeline. Full-text analysis only happens for open access material (or where you supply your own subscription plus their browser extension). Otherwise it works from title and abstract only - and extraction from an abstract is precisely how you get the last failure in the list above, the confident figure the full text qualifies. Abstracts are written to be quotable. That is the problem.
Semantic Scholar's tldr field is a genuine convenience across 214M papers, and it is machine-generated. It must never be quoted as the authors' words, and it must never be the source of a number in your table. Its useful sibling is openAccessPdf, which is the practical bridge to the thing you actually need: the full text, so you can put a page number in the location column.
General rule for this whole layer: a tool that summarises is doing a different job from a tool that extracts, and the outputs look identical in a spreadsheet cell. If you cannot tell which one produced a value, treat it as a summary and go and get the number yourself.
Three columns, three failures they catch 6 min live
Funder, date-against-mechanism, and population match. They get dropped because they feel like bureaucracy next to the number, and each one catches a category of error that no amount of care elsewhere will find. The third is the most common silent failure in applied research, and the one most likely to be in something you have already shipped.
LiveFunder and interest: not worthless, not sufficient3 min▶
A vendor-funded study is not worthless. Vendors often have the only data that exists, and pretending otherwise gets you a memo with nothing in it. The rule is narrower and more useful: a vendor-funded finding needs a non-conflicted source alongside it before it carries anything. On its own it is context. Paired with an independent source pointing the same way, it becomes support - and if the independent source points a different way, you have found the most interesting sentence in your memo.
This is also why the column has to say interest and not just funder. An unfunded study by a team who built the tool has an interest. A regulator's guidance has an interest, a legitimate and declared one. Write down who benefits if the finding is true, and the weighting decision makes itself.
LiveDate against the mechanism, not against the calendar3 min▶
"Is 2019 too old?" is an unanswerable question. The answerable version is: has the mechanism this source describes changed since it was written?
- A 2019 procurement study of how universities buy software may be perfectly current. Committee structures, budget cycles and approval chains move slowly, and a six-year-old finding about a twelve-month procurement cycle is probably still true.
- A 2019 model-capability study on transcription accuracy is archaeology. The thing it measured no longer exists in the form it measured it.
Same year, opposite verdicts. This is why your date window needed a stated reason back in b4 - the reason is a claim about which mechanisms are stable, and writing it down lets you defend a nine-year-old source and reject a two-year-old one in the same memo without looking inconsistent.
Self-studyPopulation match: the silent one3 min read▶
This is the failure that gets past everyone, because nothing about it looks wrong. The source is excellent. The design is strong, the sample is large, the number is real, the paper is peer reviewed and current. It is about a different group of people, and so it quietly answers a question you did not ask.
For the Cadence decision the boundaries are sharp and easy to cross. Higher education is not K-12 - different consent regime, different budget holder, different device environment, different everything that matters to this product. Neither of them is corporate learning, which is where most of the available note-taking evidence actually lives, because that is where the money has been. A superb study of note-taking tools in corporate training tells you something real. It does not tell you what a university procurement committee will do.
The check is mechanical and takes ten seconds per row: read your claim, read the population cell, and ask whether they are the same people. When they are not, you have three honest options - drop the source, or keep it and state the transfer explicitly ("evidence is from corporate learning; we assume partial transfer and treat this as weak support"), or go and find the on-population source. What you may not do is let it sit in the support column unmarked, because that is a claim about undergraduates resting on evidence about sales teams.
2 of 10 to 4 of 10, live ★ 12 min · everyone runs it
The lab opens where b3 left you: sharp question and source tiering on, 2 of 10 claims surviving, 22 sources in play. Turn on Extraction table and two more claims come back - and the interesting part is which two. Read them before and after. Every claim that revives at this rung is a claim that asserts a number, and every one of them was failing on the same rule: the number was there but nothing recorded where it came from.
Start at 2 of 10. Read the four claims that assert a number and note the failure rule on each - "number recorded" is doing all the work.
Predict which will revive. Two of the four have a source in the set that carries the figure; two do not. Guess before you click.
Turn on Extraction table. You are at 4 of 10. Check your prediction against the two that came back.
Now name the triage bucket. These claims failed on work, not honesty - nobody had to soften a sentence, somebody had to go and record a page number. Contrast that with the contradiction failures waiting in b7, which are fixed by wording, not by effort.
Look at the two number-claims still red. One is short of a source in the set; one is the honest unknown that never survives. Write down which is which - you will need the distinction in b10.
Try it yourself - this week ◐ 40-50 min total
- Build the extraction table for your own current question, one row per screened source, using the twelve columns. Leave cells empty rather than inferring - the empty cells are a map of what you still have to read.
- Add the three skipped columns to a table you already have: funder or interest, date-against-mechanism, population match. Mark every row where the population is not the population in your claim.
- Take the last deck you presented and verify every number in it against its source. Time yourself. That number is your argument for doing this at extraction time instead.
- Find one number in your own published or shipped work with no recorded location. Go and find it, record page and table, and notice whether the number still says what you said it said.
Official sources covered
This session is mostly craft rather than standard, so the source material is the tooling layer and its documented limits, plus the extraction discipline the evidence-synthesis standards assume you already have. This page covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Why is "location in source" the column that saves you?
Re-checkability is the difference between evidence and a rumour with a citation attached. Without a location, a figure loses a qualifier at every hop - population widens, "weekly" becomes "regular", 38% becomes "nearly half" - and it always drifts toward what the reader wants.
2 · A strong, current, peer-reviewed study of note-taking tools in corporate training is sitting in the support for a claim about university students. What is wrong?
This is the most common silent failure in applied research, because nothing about it looks wrong. Higher education is not K-12 and neither is corporate learning. Drop it, or keep it and state the transfer explicitly as weak support - what you cannot do is leave it in the support column unmarked.
3 · Where does AI extraction typically go wrong?
These are not hallucinations, which is exactly why they survive review - every value is real and checkable, just about something else. No pass that reads only your table can catch them. You have to return to the source, which is what the location column makes cheap. Note that Elicit works from title and abstract only for non-open-access material, so the abstract-versus-full-text failure is the likeliest of the five.