Series
Software engineering, checked
Fifty years of software engineering folklore, from the 100× bug to the 10× programmer, checked against the research. One change followed from idea to old age, and what survives at each step.
13 parts · 211 claims checked
Evidence map
The parts
The leprechaun hunt
Fifty years of software engineering folklore, 283 claims, one question: which of the field's famous numbers survive contact with the evidence?
9 claims checkedDecide: most ideas fail
Only a third of the ideas Microsoft tested moved their metric. What the evidence says about deciding what to build, the waterfall that never was, and agile's rituals.
19 claims checkedEstimate: the arithmetic of being late
The CHAOS report's 189%, the cone of uncertainty and the average overrun that doesn't exist. Why software is late, and the few things that make estimates better.
19 claims checkedOrganise: you ship your org chart
Conway's law, Brooks's law and the bus factor. What the research says about how the shape of a team ends up in the shape of its software.
18 claims checkedWrite: the language wars
The GitHub study that seemed to settle the language wars, and the reanalysis that unsettled it. What types, memory safety, names and indentation actually do.
20 claims checkedReview: what other eyes catch
Developers say code review is for finding bugs. The comments say otherwise. What review actually catches, the rituals that didn't hold, and what pair programming really buys.
20 claims checkedTest: the failures we could have prevented
Most catastrophic failures hide in error handlers nobody tested. What the evidence says about TDD, coverage, mutants, flaky tests, fuzzing and formal methods.
20 claims checkedShip: Fridays, outages and the CI promise
"Never deploy on a Friday" rests on a gap of less than one point. What continuous integration delivers, what speed costs, and what actually takes services down.
20 claims checkedAge: how software actually rots
Maintenance is 60–90% of the cost, smells creep in, refactoring is safe, minor versions don't break. What happens to software after it ships, checked against the research.
20 claims checkedPeople: the 28:1 legend
The 10x programmer, the camel with two humps and the value of experience. How different programmers really are, and what the evidence says actually helps them.
20 claims checkedMachines: AI, checked (version 1)
55.8% faster in the lab, 19% slower in the field. What AI coding tools have changed so far, from security to benchmarks to learning, in the part that will age fastest.
19 claims checkedEvidence: how the field fools itself
In defect prediction, who ran the study explained more than the method did. Why none of the series' 283 claims rates ~95%, and what to do with evidence this weak.
20 claims checkedWhat survives
Strip fifty years of software engineering folklore down to what survives: the life of a change on one page, three lessons, a hall of shame and where to start.
34 claims checked