This month trains you to tell the form of research from its substance — in your own retrospective series, your radiomics work, your AI-assisted reviews — and to write with enough clarity that the substance can't hide.
Quick check: Buarque de Holanda's "cordial man" describes Brazilian public life as governed chiefly by —
And the Month 07 essay asked you to name what makes building a serious skull base center in Brazil both uniquely hard and uniquely meaningful. In one phrase, what was the hard-won answer you were pointed toward?
Month 7 turned outward — toward institutions, toward Brazil, toward what you're trying to build for others. This month turns toward the discipline underneath all of that building: research. You are not a hobbyist here. You have a PhD due February 2027, a growing skull base registry, radiomics work, and AI-assisted systematic-review pipelines running in parallel — real methodological machinery, built and operated largely by you. That is exactly the position where it is easiest to produce something that *looks* like rigorous science while quietly skipping the parts that make it actually true.
Feynman's cargo cult science is not an insult aimed at fraud. It's a description of a much more common failure: doing every visible motion of research correctly — the stats, the figures, the submission — while missing the one thing that actually makes it science, which is a relentless, almost paranoid honesty about how you might be fooling yourself. Ioannidis then hands you the mechanism: small samples, flexible analyses, and a hot, crowded field are enough on their own to make most published findings false, with no dishonesty required at all. Zinsser closes the loop from an unexpected angle — muddy writing is usually muddy thinking wearing a disguise, and a clear introduction is often the first place a weak design gets caught, including by its own author.
Read that question twice this month — once about your data, once about your own finishing.
This isn't abstract for you. A retrospective surgical series with 40 EEA cases, a radiomics pipeline trained on a modest cohort, an AI-assisted SR/MA extracting from studies that themselves may not replicate — every one of these sits precisely in Ioannidis's danger zone: small samples, many possible analytic choices, a field moving fast enough that nobody has had time to fail to replicate anything yet. The discipline this month builds isn't cynicism about your own work. It's the specific, learnable skill of staying honest about it.
There's a second, quieter reading of "form vs. substance" this month insists on, and it's about your relationship to finishing rather than to data. Building the registry, the pipeline, the protocol — that is substance, and you do it exceptionally well. But a body of built infrastructure that never crosses the finish line into a submitted, reviewed, published claim is, at the level of your career and your field, itself a kind of form without substance: the machinery of research, running, producing nothing anyone else can check, cite, or build on. The texts below teach honesty about your data. The section after them asks you to turn that same honesty on your own finishing.
It's worth naming why the machinery keeps growing while the pipeline of finished work doesn't: building a new extraction schema or a tighter de-identification protocol has urgency built in — a new dataset to process, a novel pipeline to test, an audience of one (you) to satisfy immediately. A finished submission has none of that pull. It sits with an unknown editor, an uncertain timeline, and a verdict that arrives weeks or months later, disconnected from the moment you did the work. Interest and stakes drive your activation far more reliably than importance does — the same pattern this course names for the OR versus the paperwork shows up here as the registry versus the submission.
How to read it: one sitting — it's short, a commencement-style address, meant to be read the way it was delivered: with warmth and a bit of mischief, not as a stern methods lecture. Read it once for the story (the Pacific islanders building runways and control towers out of straw, waiting for planes that never land), then a second time slower, marking every sentence where he names a *specific* way scientists fool themselves.
The core move: the islanders copied every visible form of an airport — the runway, the headphones, the antenna made of bamboo — without the one thing that actually makes planes land, which is the whole invisible infrastructure of radar, radio, and real air traffic behind it. Feynman's claim is that plenty of published science does the same thing: it has the shape of a paper (methods, results, p-values, a discussion section) without the substance, which he defines as an obsessive, almost self-punishing honesty — reporting *everything* that might be wrong with your own result, including your own data selection, your own biases, and the experiments that failed.
Your radiomics pipeline reports strong predictive performance on your training cohort. Before you call this a finding, which move is Feynman actually asking for?
A second scenario, closer to home: a PROSPERO registration sits in "draft" for six weeks while you keep refining the search string, telling yourself the protocol isn't "ready" yet. Feynman would call this —
How to read it: one sitting for the argument, then keep it as a reference you re-read before any extraction or analysis decision on your own studies. Read it as rigor, not despair — Ioannidis is not saying science is broken; he is naming, with a formula, exactly *which* research conditions make false-positive findings statistically likely, so you can recognize when you're operating inside them.
The core move: a published finding is more likely to be false when studies are small, effect sizes are small, there are many things being tested (many possible outcomes, subgroups, or analytic paths), there is flexibility in study design and analysis, financial or intellectual interests are large, and — critically for you — the field is "hot," meaning many teams are chasing the same question at once. None of these require misconduct. They are structural conditions that make a false positive the statistically expected outcome, not the exception.
One more idea worth carrying: Ioannidis distinguishes a study's statistical power from its positive predictive value — the probability that a "significant" finding is actually true, given how rare true effects are in the field you're sampling from. In a hot, crowded field chasing a genuinely rare true effect, even a well-powered single study will more often be wrong than right, purely from base rates. This is the argument for treating any single striking result — yours or someone else's — as a hypothesis to replicate, not a conclusion to build on, no matter how clean the p-value looks.
Practically, this reframes flexible analysis as the enemy, not a convenience. Pre-specifying your primary outcome and analysis plan before looking at the data — even informally, in your own registry protocol — is the single cheapest defense Ioannidis's own framework implies.
There is a second, quieter implication for you specifically: Ioannidis's framework doesn't just grade the finding, it grades the incentive structure around finishing. A registered protocol with a null result is boring — no urgency, no audience, no novelty, exactly the conditions under which, per M03/M08's shared pattern, your own activation goes quiet. The paper most likely to sit unfinished is not the flashy positive result; it's the honest, modest, pre-specified one that confirms the design worked but found nothing exciting. That is precisely the paper Ioannidis says the literature needs more of — and precisely the one your wiring is least drawn to finish.
Hold that thought — section 07 below names this directly and asks for a matching action, not just a matching insight.
Your 40-case retrospective series shows a numerically lower complication rate with a modified approach. Which claim is actually licensed by the design?
Your AI-assisted SR/MA pipeline lets you re-run the meta-analysis with different inclusion cutoffs in minutes. After the fourth re-run, one cutoff finally produces a significant pooled estimate. What does Ioannidis's framework say about this specific effect?
One reference idea worth carrying alongside Ioannidis, from Austin Bradford Hill's 1965 address "The Environment and Disease: Association or Causation?": a short set of considerations — strength of association, consistency across settings, a plausible mechanism, a coherent dose-response pattern, temporality (the cause must precede the effect) — that help you judge whether an association in your own surgical series is worth treating as causal, or should stay labeled as an association. Hill never intended these as a checklist to "pass" a finding into causal language; he intended them as a discipline for resisting overclaiming from observational data — exactly the discipline Ioannidis's math demands and Feynman's honesty requires.
How to use it: not as homework to complete before every claim, but as a five-second internal audit whenever a sentence in your discussion section starts drifting toward causal verbs ("reduces," "prevents," "improves") instead of associative ones ("is associated with," "was lower in," "was consistent with"). If you can't answer at least strength, consistency, and temporality with something concrete, the sentence needs a softer verb, not a longer methods section.
For your retrospective series specifically: a single small cohort showing a pattern, with a plausible surgical mechanism and consistency with prior literature, is worth reporting as a hypothesis-generating association. It is not, on its own, evidence of causation — that gap is exactly where "form of research" quietly slides into "substance."
Two of Hill's nine considerations deserve special weight in exactly your kind of work. Temporality is non-negotiable — the proposed cause must be shown to precede the effect, which sounds obvious until you notice how easy it is, in a retrospective chart review, to reconstruct a sequence that fits the conclusion you already suspect. Consistency — does the pattern hold across your own cases and the handful of other published series — is the cheapest check available to you before claiming anything, and the one most often skipped when a result is exciting enough to want to publish quickly.
Your registry shows a plausible mechanism (a technical modification reduces manipulation of a specific structure) and a pattern consistent with two small published series. Strength of association is modest and the cohort is retrospective. Is this ready for causal language?
How to read it: read the early chapters on clarity and clutter closely — those are the ones that change how you write manuscripts and grants; skim the later genre-specific chapters (memoir, sports, humor) unless one speaks to you directly. This is not a style manual to memorize; it's a set of habits to apply immediately to your next paragraph.
The core move: clear writing is not decoration on top of clear thinking — it is clear thinking, made visible enough that a reader (or you, six months later) can find the flaw. Zinsser's central discipline is clutter-hunting: every sentence has words doing no work — hedges, throat-clearing phrases, unnecessary qualifiers, passive constructions that hide who did what. Strip them, and what's left is either a sound idea stated plainly, or a weak idea suddenly visible as weak. Muddy prose is frequently a weak argument's best hiding place — including from its own author.
A second idea from the same chapters: Zinsser insists a writer should picture one real reader, not an imagined committee of critics — write the sentence you'd say out loud to that one person, then trim it to match. This is a direct antidote to a specific trap: writing for the imagined harshest possible reviewer produces defensive, over-hedged prose exactly because you're pre-negotiating with a jury in your head instead of stating what you found to one honest reader.
This matters for exactly the documents your life runs on: manuscript introductions (where a fuzzy justification for the study conceals a fuzzy rationale for doing it), grant applications (where clutter reads as either padding or evasion to a reviewer with eleven other applications to read that day), and patient materials (where clarity is not a courtesy — it is the difference between informed and merely signed consent). And, not incidentally, this is the same skill your international academic ambitions depend on: a clear English abstract travels; a cluttered one doesn't get read past the first paragraph.
One more habit from Zinsser worth adopting mechanically: read your own paragraph aloud before sending it. Clutter that survives silent proofreading rarely survives being spoken — the ear catches the stacked hedge, the passive construction, the sentence that takes three tries to say, in a way the eye skims past. This costs ninety seconds per paragraph and catches a category of error no spell-checker or reference manager will ever flag.
A manuscript introduction opens: "It has been increasingly recognized in recent years that the management of skull base tumors, while historically challenging, may potentially benefit from a more nuanced consideration of approach selection in select patient populations." What is Zinsser's diagnosis?
A surgical consent explanation reads: "There is a possibility that certain complications, while relatively uncommon, may potentially occur in association with this type of procedure." A patient signs it without asking a single question. What has actually happened?
A grant application's aims page reads: "We seek to comprehensively explore, characterize, and further elucidate the multifaceted role of imaging biomarkers." What would Zinsser cut first?
Four short research moments from your world. Tag each: is this the form of rigor (looks right, isn't) or the substance (actually honest, actually earns the claim)?
A radiomics paper reports the single best-performing model after testing 30 feature combinations, without mentioning the other 29.
A pre-registered protocol states the primary outcome and analysis plan before the SR/MA extraction begins, and the manuscript reports it even when the result is null.
A 40-case surgical series is titled and discussed using language that implies the approach is now "validated" as superior.
A discussion section explicitly names the study's small sample, single-center bias, and the specific ways the finding could still be wrong.
A manuscript has been marked "FINAL" in its filename for two months while its PROSPERO registration sits in draft, unregistered.
An AI-assisted extraction pipeline flags every disagreement between the model and a human reviewer for manual adjudication, and the adjudication log is versioned alongside the dataset.
The form is often invisible to the author precisely because it uses all the right vocabulary — p-values, citations, hedged language. The tell is always what's *missing*: the disconfirming path, the honest limitation, the pre-specified plan — or, in your case specifically, the submission itself.
Here is the number this month is actually about, and it isn't a p-value: 373 registry cases, zero approved manuscripts. Not one dataset assembled with less rigor than that would demand — the opposite. Every one of those cases sits inside a registry protocol, a de-identification pass, an extraction schema, verification gates you built and enforce on yourself. The craft is real. Ioannidis's disclosure discipline, Feynman's self-directed honesty, Zinsser's clutter-hunt — you already run all three, upstream, better than most of the literature you critique. The 90% is not where this month's danger lives.
It lives in the last 10%: the PROSPERO submission, the "Submit" button on the manuscript portal, the reviewer response you've been "about to" send. Manuscripts sit FINAL with PROSPERO still pending — the writing is done; the registration, which takes an afternoon, is not. This is the shape the Pact dossier calls the perfectionism paradox: your surgeon's craft rigor shows up everywhere except at the finish line, because shipping is an evaluative moment and building is not. A draft can always get one more pass and nobody judges you for it. A submission gets a verdict — accept, reject, revise — and higher standards make that verdict scarier, not safer, which is exactly backwards from what you'd expect from someone this rigorous.
Notice what this month's texts have been quietly training you to see: Feynman's cargo-cult science is a paper that has the form of research without the substance of finishing the honest reckoning. Your version runs in the opposite direction — you have done the substance (the registry protocol, the extraction, the honest limitations section) and stalled at the form: the actual, external, checkable act of submitting it. Both are avoidance that photographs like virtue. One postpones honesty by skipping it; yours postpones the verdict by perfecting around it, indefinitely, one more time.
There is no trick for this that skips the discomfort — the fix Ioannidis's own framework points to is structural, not psychological: a pre-specified stopping rule. Not "when it feels ready" (a feeling you can manufacture doubt about forever) but a date, witnessed by someone else, after which the current draft ships regardless of how it feels. This is the same move as a pre-registered analysis plan applied to your own behavior: decide the criterion for "done" before you're inside the anxiety of the decision, because inside it, "one more pass" will always sound reasonable.
A colleague's SR/MA has been "final, just needs one more read-through" for three weeks. She reads it again tonight and finds two more sentences to tighten. What is actually happening?
Read that scenario back onto your own registry and manuscripts. The tell is the same one the dossier names: a body of built assets that keeps compounding, next to a submission count that doesn't move. The craft was never the gap. The finish line was.
373 registry cases and zero approved manuscripts is not a verdict on the science. It is a verdict on the ratio of building to shipping — and only one of those two numbers is what this month is asking you to move.
Say plainly what this is not: it is not a case for lowering the bar. Your verification gates, your insistence that pilot-N evidence gets called "pilot," your refusal to let a manuscript claim more than the design earns — none of that should soften. In a physician-researcher, that rigor is a compounding asset; it is the reason your name on a paper will eventually mean something specific. The paradox is precisely that this same rigor, aimed at the wrong target — the draft instead of the deadline — becomes the mechanism that keeps the paper from ever reaching a reader who could benefit from it. Feynman never asked for an unfinished, honest paper. Ioannidis never asked for a null result that stays in a drawer. The honesty this month asks for has to include honesty about when "not yet" has become the actual answer, indefinitely.
A second angle on the same fresh scenario: your colleague could respond to feedback about her three-week stall by resolving to "be more disciplined about writing time." Would that fix it?
A third variant: you finally book a "finishing session" for a stalled PROSPERO registration, but when the day arrives you spend the whole hour reorganizing the search-string spreadsheet instead of opening the registry portal. What actually happened?
This is why the practice below asks for a rewritten introduction and a real submission-adjacent action this month — not because the writing exercise isn't valuable, but because a formation note that only ever touches the writing, never the shipping, would itself be the pattern this course exists to break.
Name the strengths too, since this month is easy to misread as an indictment: the willingness to kill sunk costs (MedOps, Clawd, Synthesium, the CCF papers you chose not to force into research-lab) is the same rigor working correctly in the other direction — recognizing when something doesn't earn its continuation. The finish line asks for that same clear-eyed judgment turned toward what's ready, not just what isn't worth keeping.
Before the practice, sit with the texts themselves for a moment. These five passages are the load-bearing walls of the month — each one paraphrased tightly, each one aimed at a specific decision you will face in your own research this year. Read them slowly; the "em termos simples" line under each is the version to carry into the next analysis session.
His closing principle, the one everything else in the address serves: "the first principle is that you must not fool yourself" — and, he adds, you are the person easiest for you to deceive. Integrity in science, for Feynman, is not avoiding lies to others; it is a deliberate, effortful bending-over-backwards to expose to yourself every way your own result could be wrong, before anyone else gets the chance.
Em termos simples: before any radiomics or registry result leaves your machine, write down the three ways it could be an artifact of your own choices — cohort, features, cutoffs — and only then decide if it survives.
Beyond honesty with yourself, Feynman demands honesty in what you publish: report every detail that could cast doubt on your interpretation, name the alternative explanations you ruled out and how, and give readers enough information to judge for themselves — not just the information that leads to judgment in one particular direction. Selective reporting, even of true facts, is for him already the cargo cult.
Em termos simples: in your SR/MA and series manuscripts, the limitations section is not defensive boilerplate — it is the part of the paper where the science actually happens; write it first, not last.
The paper's engine is a single quantity: the post-study probability that a claimed finding is true, which depends less on the p-value than on the prior odds that a true effect exists, the study's power, and — his sharpest addition — bias and the number of teams chasing the same question. His six corollaries all point one way: small studies, small effects, many tested relationships, flexible designs, financial interest, and hot fields each independently lower the chance a positive claim is real. Under realistic values, he argues, "most claimed research findings are false" — with no fraud required anywhere in the chain.
Em termos simples: your 40-case series and your radiomics cohort tick four of his six risk corollaries before you run a single test — so pre-specify the primary analysis and label everything else exploratory, every time.
Hill lists nine considerations for judging whether an observed association is causal — strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, analogy — and then, in the move most readers miss, refuses to make them a checklist: none, except temporality, is necessary; none is sufficient; they are aids to judgment, not a scoring rubric. He closes on humility — "all scientific work is incomplete" — yet insists that incompleteness never excuses postponing the action the evidence already justifies.
Em termos simples: when a discussion-section sentence wants a causal verb, run strength, consistency, and temporality against your registry data — and if any of the three is missing, downgrade the verb, not the standard.
Zinsser's diagnosis, from the opening chapters: "clutter is the disease of American writing" — the stacked qualifiers, the pompous frills, the sentence that could not survive being said aloud. His prescription is surgical, not cosmetic: strip every sentence to its cleanest components, and treat the discipline of simplifying as identical to the discipline of thinking, because a writer who cannot say a thing plainly usually does not yet know what the thing is.
Em termos simples: the hedge-hunt in this month's practice is not a style pass — every "may potentially" you delete from a manuscript forces you to decide what you actually claim, which is the research decision itself.
Caro recounts the instruction from his first newspaper editor that became his life's method: "turn every page" — never assume anything, never skip the file that probably holds nothing, because the truth of how power actually worked was always in the document nobody had bothered to read. His entire book is a defense of unglamorous thoroughness: the years in the Lyndon Johnson archives, the patience to wait as long as the truth takes.
Em termos simples: in an SR/MA, "turn every page" means the full-text you almost excluded on the abstract, and the supplementary table your AI pipeline skimmed — the finding that changes the forest plot is usually hiding there.
Loci: Feynman, closing paragraphs of the 1974 Caltech commencement address · Ioannidis, PLoS Med 2(8):e124, corollaries 1–6 · Hill, Proc R Soc Med 58:295–300, the nine considerations and closing section · Zinsser, On Writing Well, chs. 2–3 · Caro, Working, the "turn every page" chapter on his Newsday apprenticeship.
Take one manuscript introduction you've written (or are drafting) and rewrite it applying Zinsser's clutter-hunting and Ioannidis's honesty about what the design can claim.
Exact steps:
learning-records/writing-craft/ as YYYY-MM-DD-intro-rewrite.md, before-and-after, with one line naming what the "before" version was hiding.Manuscript / section: ___ BEFORE (verbatim, current draft): ___ Hedges and clutter I circled: ___ AFTER (rewritten — gap / what we did / honest claim, 3 sentences max): ___ What the "before" version was quietly hiding or overclaiming: ___
"It has been increasingly recognized that approach selection may potentially benefit from more nuanced consideration in select populations."
"Approach selection for this tumor location lacks direct comparative data. We reviewed 40 consecutive cases to describe outcomes by approach. Given the sample size and retrospective design, these findings are hypothesis-generating, not comparative evidence."
"Done" looks like: one before/after file where the "after" version makes a smaller, honest claim that survives being read aloud — not a bigger claim dressed in plainer words.
learning-records/writing-craft/ exists and one real submission-adjacent action moved forward this month — a PROSPERO registered, a manuscript actually submitted, a reviewer response actually sent. A rewritten intro that stays on your laptop is collection, not formation; the whole point of this month is that the writing was never the bottleneck. If nothing external moved, the practice isn't done, no matter how good the "after" paragraph reads.A note without a changed behavior is a mirror you admired, not formation — this course is itself a mirror, and it only counts if it converts.
This month sharpens the morning line for research days: before opening a manuscript, registry export, or analysis notebook, name one honest question — "what result am I hoping to find today, and would I trust it if it came from someone else's small, hot-field, flexible-analysis study?" It also feeds the evening five minutes on writing days: did today's draft state a smaller true thing, or a bigger unearned one?
The dichotomy-of-control checklist from Month 1 applies directly here: sample size, field crowdedness, and prior literature quality are largely not up to you; pre-specifying your analysis, disclosing your limitations, and writing without clutter are entirely up to you. Put your effort where it actually changes the outcome.
One addition specific to this month's practice: schedule the actual finishing act — PROSPERO submission, manuscript upload, reviewer response — the same way you schedule the OR, on the calendar, at a specific time, ideally with someone who will ask whether it happened (Samilly, a co-author, a lab meeting). This is the finishing-session move the Pact protocol already uses elsewhere: staged stakes and a witness turn an infinitely postponable act into one with a deadline and an audience, which is exactly what your activation pattern responds to. You are not fixing the anxiety about the verdict; you are borrowing urgency from a source other than the verdict itself.
A cadence target worth writing into the protocol rather than leaving implicit: roughly one submission every six to eight weeks, each one small enough to be drivable by you alone without waiting on a co-author's calendar. Not every submission needs to be the definitive paper on a topic — a well-labeled pilot series, a registered protocol, a short correspondence piece all count. The goal this cadence protects against is the one Ioannidis's own framework would predict from your specific incentive structure: an ever-growing backlog of high-quality, unfinished substance, indistinguishable from no research program at all to anyone outside your own head.
Claim only what the design earns, and say it plainly.
The paper is a case report, not a pivotal trial. Ship it like one.
Feynman and Ioannidis, and Zinsser's editor at your shoulder, would all sign off on the same corollary: honesty about the data means nothing to a reader who never receives the paper.
A closing distinction worth keeping, since it's easy to blur the two under one word ("rigor"): data honesty (Feynman, Ioannidis, Hill) protects the reader from a false claim; finishing discipline (this month's sob-medida) protects the reader from never getting a claim at all. You have built the first in abundance. This month's actual test is the second.
Bring it back to me. After your first before/after rewrite, tell me the one sentence that changed the most — the one where clarity forced you to admit the claim was smaller than you wanted it to be. That's the tell the practice is working.
And bring back the other half. Tell me the date you set for the finishing act, and whether it happened. If it slipped, that's not a failure to hide — it's the exact data point this month exists to surface: name what made "one more pass" feel necessary in the moment, and whether the witness actually got asked.
This lesson isn't done when you finish reading — it's done when one intro-rewrite.md exists in learning-records/writing-craft/ and one real submission moved, however small.
Primary texts: Feynman, "Cargo Cult Science" · Ioannidis, "Why Most Published Research Findings Are False," PLoS Med (2005) · Zinsser, On Writing Well. Supporting: Bradford Hill (1965). Full rationale in SYLLABUS.md; shelf notes in RESOURCES.md.
This month is done at the surgeon's definition, not the reader's: not "I read the three texts," but "one intro-rewrite exists in learning-records/writing-craft/, and one submission-adjacent act — PROSPERO, upload, or reviewer response — actually happened, witnessed."