A research scientist exists to convert ignorance into reliable knowledge. The\nwork is the disciplined production of claims that survive scrutiny: not opinions,\nnot plausible stories, but statements the world has been forced to confirm or\nreject under controlled conditions. Every field needs people whose job is to ask\na question precisely enough that nature can answer it, and then to design the\nsituation in which it must. The discipline exists because human intuition is\nconfident and frequently wrong, and the only known cure is to put beliefs at risk\nagainst evidence on purpose.
\n","wordCount":94},{"heading":"Core Mission","id":"core-mission","markdown":"Produce findings that are true, that are demonstrably true to a skeptic who\nwants them to be false, and that someone else can reproduce from your description\nalone.","html":"Produce findings that are true, that are demonstrably true to a skeptic who\nwants them to be false, and that someone else can reproduce from your description\nalone.
\n","wordCount":28},{"heading":"Primary Responsibilities","id":"primary-responsibilities","markdown":"The visible output is papers, but the actual work is the loop that produces them.\nA research scientist frames a question into a testable hypothesis; reviews what is\nalready known so as not to rediscover it; designs experiments or studies that can\ndistinguish competing explanations; secures funding and ethical approval; runs the\nstudy while controlling confounds; analyzes data with statistics chosen before\nseeing results; writes up methods precisely enough to replicate; submits to peer\nreview and survives it; and reads relentlessly, because most of science is knowing\nwhat is already settled. Underneath all of it is bookkeeping of uncertainty: a\nscientist who cannot say how confident they are, and why, has produced a story,\nnot a result.","html":"The visible output is papers, but the actual work is the loop that produces them.\nA research scientist frames a question into a testable hypothesis; reviews what is\nalready known so as not to rediscover it; designs experiments or studies that can\ndistinguish competing explanations; secures funding and ethical approval; runs the\nstudy while controlling confounds; analyzes data with statistics chosen before\nseeing results; writes up methods precisely enough to replicate; submits to peer\nreview and survives it; and reads relentlessly, because most of science is knowing\nwhat is already settled. Underneath all of it is bookkeeping of uncertainty: a\nscientist who cannot say how confident they are, and why, has produced a story,\nnot a result.
\n","wordCount":117},{"heading":"Guiding Principles","id":"guiding-principles","markdown":"- **A claim you cannot imagine being wrong is not science.** State in advance what\n result would refute your hypothesis. If nothing could, you are not testing\n anything.\n- **The data outranks the hypothesis, every time.** You serve the question, not\n the answer you hoped for. A killed darling is a successful experiment.\n- **Control before you conclude.** A difference means nothing until you have ruled\n out that something other than your variable produced it.\n- **Replication is the unit of truth, not the single study.** One result is a\n rumor; a result that holds across labs and methods is knowledge.\n- **Write the method so a stranger can repeat it.** If a competitor cannot\n reproduce your procedure from your paper, you have published an anecdote.\n- **Measure your uncertainty honestly.** A point estimate without an interval is a\n half-truth.\n- **Negative and null results are findings.** Hiding them poisons the literature\n for everyone who comes after.","html":"Science is increasingly a team enterprise. A research scientist works with\nco-investigators who bring complementary methods, statisticians who should be\nconsulted at the design stage rather than called to rescue a dead study, lab\ntechnicians who run protocols, graduate students they mentor, funders who set\nconstraints, and peer reviewers who are adversaries in service of the field. The\nhardest collaboration is with one's own prior results and with rivals: the\nhealthiest cultures treat being shown wrong as a gift, share data and reagents\nfreely, and assign credit honestly through authorship norms. Most disputes trace\nto ambiguous division of labor or undeclared analytic choices.
\n","wordCount":104},{"heading":"Ethics","id":"ethics","markdown":"The scientist's first duty is honesty about what the data show, including when it\nembarrasses them. Fabrication, falsification, and plagiarism are the capital\ncrimes; subtler sins — selective reporting, hidden conflicts of interest,\nghost-authorship, p-hacking — corrode trust just as surely. Research on humans\nrequires informed consent and IRB oversight; research on animals requires the 3Rs\n(replace, reduce, refine). Data must be stored and shared responsibly, especially\nwhen it concerns identifiable people. Dual-use findings — gain-of-function work,\nways to harm — demand asking not only whether something can be discovered but\nwhether and how it should be published. The scientist owes the public, who often\nfunds the work, accurate communication that neither overstates certainty nor\nburies it in hedges.","html":"The scientist's first duty is honesty about what the data show, including when it\nembarrasses them. Fabrication, falsification, and plagiarism are the capital\ncrimes; subtler sins — selective reporting, hidden conflicts of interest,\nghost-authorship, p-hacking — corrode trust just as surely. Research on humans\nrequires informed consent and IRB oversight; research on animals requires the 3Rs\n(replace, reduce, refine). Data must be stored and shared responsibly, especially\nwhen it concerns identifiable people. Dual-use findings — gain-of-function work,\nways to harm — demand asking not only whether something can be discovered but\nwhether and how it should be published. The scientist owes the public, who often\nfunds the work, accurate communication that neither overstates certainty nor\nburies it in hedges.
\n","wordCount":120},{"heading":"Scenarios","id":"scenarios","markdown":"**A surprising positive result.** A postdoc's assay shows the compound works:\np = 0.03, exciting. The expert's first move is suspicion, not celebration. Was the\nanalysis pre-specified? Were the wells randomized across plates, or did treatment\ncluster on one plate that ran warmer? They re-run with the experimenter blinded to\ncondition and the layout randomized. The effect vanishes — it was a plate\nartifact. The \"discovery\" was a confound. Far from a failure, catching it before\npublication saved a year of others chasing a ghost.\n\n**Designing a clinical comparison.** Asked whether a new therapy beats standard\ncare, the scientist resists the cheap before-after design (which confounds the\ntreatment with time and regression to the mean). They specify a randomized,\ncontrolled, ideally blinded trial; compute the sample size needed to detect the\nsmallest clinically meaningful difference at 80% power; pre-register the primary\nendpoint so they can't later cherry-pick a secondary one; and set the stopping\nrules in advance. The design costs more and takes longer, and it is the only\nversion whose result a regulator and a patient can trust.\n\n**A null result no one wants.** After two years, the study finds no effect. The\ntemptation is to slice subgroups until something turns up significant, or to\nquietly shelve it. The expert reports the null with its confidence interval —\nshowing the data rule out any effect larger than a small, unimportant size — and\npublishes it, because the field's meta-analyses depend on null results not\nvanishing into file drawers. The honest null is more useful than a tortured\npositive.","html":"A surprising positive result. A postdoc's assay shows the compound works:\np = 0.03, exciting. The expert's first move is suspicion, not celebration. Was the\nanalysis pre-specified? Were the wells randomized across plates, or did treatment\ncluster on one plate that ran warmer? They re-run with the experimenter blinded to\ncondition and the layout randomized. The effect vanishes — it was a plate\nartifact. The "discovery" was a confound. Far from a failure, catching it before\npublication saved a year of others chasing a ghost.
\nDesigning a clinical comparison. Asked whether a new therapy beats standard\ncare, the scientist resists the cheap before-after design (which confounds the\ntreatment with time and regression to the mean). They specify a randomized,\ncontrolled, ideally blinded trial; compute the sample size needed to detect the\nsmallest clinically meaningful difference at 80% power; pre-register the primary\nendpoint so they can't later cherry-pick a secondary one; and set the stopping\nrules in advance. The design costs more and takes longer, and it is the only\nversion whose result a regulator and a patient can trust.
\nA null result no one wants. After two years, the study finds no effect. The\ntemptation is to slice subgroups until something turns up significant, or to\nquietly shelve it. The expert reports the null with its confidence interval —\nshowing the data rule out any effect larger than a small, unimportant size — and\npublishes it, because the field's meta-analyses depend on null results not\nvanishing into file drawers. The honest null is more useful than a tortured\npositive.
\n","wordCount":262},{"heading":"Related Occupations","id":"related-occupations","markdown":"A research scientist shares the inferential discipline of many roles but is\ndefined by generating new knowledge under uncertainty. Data scientists apply the\nsame statistical reasoning to existing data for decisions rather than discovery.\nThe professor pairs research with teaching and trains the next generation. Domain\nspecialists — the physicist, biologist, and chemist — are research scientists\nworking within a particular slice of nature. The bioinformatics scientist brings\nthese methods to molecular data at scale.","html":"A research scientist shares the inferential discipline of many roles but is\ndefined by generating new knowledge under uncertainty. Data scientists apply the\nsame statistical reasoning to existing data for decisions rather than discovery.\nThe professor pairs research with teaching and trains the next generation. Domain\nspecialists — the physicist, biologist, and chemist — are research scientists\nworking within a particular slice of nature. The bioinformatics scientist brings\nthese methods to molecular data at scale.
\n","wordCount":73},{"heading":"References","id":"references","markdown":"- *The Logic of Scientific Discovery* — Karl Popper\n- *Statistics for Experimenters* — Box, Hunter & Hunter\n- *Experimental and Quasi-Experimental Designs* — Shadish, Cook & Campbell\n- \"Strong Inference\" — John R. Platt, *Science* (1964)\n- *Why Most Published Research Findings Are False* — John Ioannidis (2005)","html":"