Ask what work the fact does
Your recent writing discussions keep returning to a useful question:
What work is this fact doing?
A fact can do any of these jobs:
- establish scale
- prove existence
- compare alternatives
- identify a mechanism
- provide a counterexample
- set a boundary
- explain chronology
- show change over time
- establish credibility
- give the reader a concrete picture
A fact with no job is research residue.
Numbers need context
A number is rarely meaningful by itself. Ask for:
- denominator
- baseline
- time period
- comparison group
- nominal versus real values, when relevant
- percentage versus percentage-point change
- sample size
- how the number was measured
- whether the source has an incentive in how the number is framed
Replace "huge," "rapid," "rare," and "significant" with numbers when the numbers exist. Do not use a number merely to create authority.
Evidence should sit near the claim it supports
Long separations between claim and evidence increase the reader's memory load. The reader should not have to remember a statistic from three pages earlier and infer its role now.
If evidence supports several later claims, either signpost that role clearly or repeat the relevant datum with restraint.
Triangulate consequential claims
Not every sentence needs two sources. Consequential, disputed, surprising, or easily misreported claims deserve stronger verification.
A useful hierarchy runs from strongest to weakest:
- A direct record, primary document, original dataset, or firsthand observation.
- A named source with direct knowledge.
- Credible secondary analysis with transparent sourcing.
- Unattributed or tertiary summaries, used mainly to locate stronger evidence.
The hierarchy changes by field, but the principle survives. When the claim matters, move closer to the event or the data.
Research must eventually stop
Zinsser's warning is practical. More research can become another form of avoidance.
Stop when any of these is true:
- The core claim and countercase are supported well enough for the intended piece.
- Additional sources repeat the model rather than change it.
- The remaining uncertainty is better disclosed than endlessly chased.
- New material keeps expanding the scope instead of strengthening the argument.
Then write. Return to research only when the draft reveals a real gap.
Use different evidence for different jobs
The Howard Marks analysis makes a useful distinction explicit. Evidence becomes richer when you do not treat its forms as interchangeable.
- Current data can establish magnitude or direction.
- Historical cases can establish precedent or reveal changing conditions.
- Primary records can establish what actually happened.
- Expert interpretation can explain a mechanism or contested meaning.
- Personal observation can establish what the writer directly experienced.
- Analogy can aid comprehension, but it cannot prove the claim.
A strong evidence mix is not variety for its own sake. Each source type should do the job it is suited to do.
Quotations are participants, not ornaments
A quotation should enter because another mind is needed at that point in the argument. Before you use one, know its role. A quotation can:
- support a claim with direct knowledge
- state the strongest opposing view
- supply a distinction or mechanism
- capture wording whose exact form matters
- show how a participant understood the event at the time
Introduce why the source matters, use only the necessary words, and return immediately to your own reasoning. A famous person saying something similar to you is not evidence merely because the name is famous.
Quantitative claims need a separate scepticism pass
Your science-writing highlights expose a gap in the first version of this course. Numbers need not only context but methodological scrutiny.
For consequential quantitative claims, ask:
- What was actually measured? Is it the outcome readers care about, or only a proxy?
- What is the denominator, baseline, comparison group, and time period?
- Is the reported change absolute or relative?
- How large is the effect in practical terms, not merely whether it crossed a statistical threshold?
- How many outcomes, subgroups, or analyses were examined before this result was selected?
- Was the hypothesis or analysis planned before seeing the data, or constructed afterward?
- Does a subgroup result have independent confirmation, or is it exploratory?
- What observations were excluded, and why?
- Who funded, produced, or publicised the analysis, and what incentives do they have?
- Has an independent expert examined the method rather than merely the headline result?
Terms such as p-hacking and HARKing describe ways analytical flexibility can create apparently persuasive findings. You do not need to diagnose misconduct. You need enough statistical scepticism to ask two things: whether the reported result was one of many possible analyses, and whether the story was chosen after the data were known.
Likewise, a p-value near a conventional threshold is not proof that anything improper occurred. Treat suspicious patterns as prompts for better questions, not as verdicts.
Write risk so the reader can form their own judgment
Risk can be framed in mathematically equivalent ways that feel psychologically different. When the distinction matters, give both absolute and relative quantities.
For example, "risk doubled" is incomplete. The reader needs to know whether it moved from 1 in 10 to 2 in 10, or from 1 in 10,000 to 2 in 10,000.
Useful risk writing usually supplies:
- baseline risk
- changed risk
- time horizon
- population to which the estimate applies
- uncertainty or confidence range, when available
- what action, if any, can materially change the odds
Avoid "prove" when the design supports association, probability, or bounded inference rather than certainty. The reader should be able to see what the study establishes and what remains unknown.
Follow the incentives behind the evidence
Do not confuse professional presentation with independent evidence. Company white papers, press releases, consultancy reports, advocacy research, and even peer-reviewed papers can have incentives that shape what is measured, highlighted, or omitted.
"Follow the money" is a starting question, not a dismissal. Funding does not automatically invalidate evidence. It tells you what deserves additional scrutiny and independent checking.