Evidence over anecdote

What sounds right is the enemy

The most dangerous sentence in analysis is the one that sounds right. A claim that fits the reader's intuition, flows well, and confirms the prevailing mood will pass unchallenged precisely because nothing about it triggers resistance — and that frictionless plausibility is exactly what lets a wrong idea travel. The anecdote is the purest form of this hazard: a single vivid story, emotionally legible, generalized into a law. Evidence over anecdote is the discipline of distrusting plausibility itself — of refusing to let a claim stand because it sounds right, and demanding instead that it be anchored in something that was actually measured: a trial, a dataset, a real deployment, a number with a traceable origin. The standard is not "is this persuasive?" but "what would have to be true for this to be false, and has anyone checked?"

This is why ENSI's strongest pieces are built from specific, checkable instances rather than evocative generalities. AI-driven eGovernment: the opportunities does not assert that AI helps government; it points at the UK's Extract tool turning scanned planning records into structured data in seconds, at Singapore's CORENET X moving to machine-readable building-code submissions, at France's tax authority scaling computer vision over aerial imagery, at Ukraine's ProZorro pairing open contracting with watchdog analytics. Each is a real deployment with a measurable result, and the argument is the sum of the deployments rather than a feeling about them. Strip the instances out and the article evaporates into the kind of optimistic vapor that fills most writing on the subject. The evidence is not decoration on the claim; the evidence is the claim, and the prose merely arranges it.

Measured, not asserted

The method draws a sharp line between three things that are routinely confused: what was measured, what was inferred, and what merely sounds right. A measured fact has a source and, ideally, corroboration; an inferred claim has a mechanism connecting it to measured facts; a plausible assertion has neither and is treated as a rumor until it earns better. Impact of generative AI on worker productivity is disciplined precisely because it stays anchored to what trials and studies have actually found about productivity, rather than reaching for the larger, more exciting, less supported claim that the topic invites. The economics of infinite intelligence makes a sweeping argument, but it rests on an observable fact about cost curves rather than on a mood about technology — the sweep is earned by the anchor. The method's instinct is to locate the measured floor under any big claim and to write only as far above that floor as the floor will bear weight.

This is also why ENSI treats meta-questions about evidence as first-class subjects rather than housekeeping. Meta science: the systemic failures of the scientific machine is, at bottom, an extended argument about how the institutions that are supposed to produce reliable evidence have developed pathologies that let plausible-but-unreplicated claims masquerade as knowledge — which is the same hazard the method guards against at the scale of a single sentence. Epistemology: from discovery to justification names the exact distinction the method enforces: having an idea is cheap, and being entitled to assert it is expensive, and the gap between the two is filled by evidence. A method that cared only about generating insight would skip that gap; ENSI's method treats the gap as the work.

Diversify toward the dissent

The subtle failure of evidence-led writing is not making things up; it is gathering only the evidence that agrees. An author who already believes a thesis can assemble a wall of confirming citations and call it rigor, when in fact they have simply collected echoes. The method counters this with a rule that feels backwards: when the evidence mostly agrees, go hunt for the strongest disconfirming case, because a thesis that has not met its best counter-evidence has not been tested, only decorated. This is why ENSI's surveys of contested terrain do not read as advocacy. National AI strategies: the common areas earns its synthesis by drawing on what many actual national strategies contain, including where they diverge, rather than projecting one preferred model onto all of them. The discipline is to treat the absence of captured dissent as a defect in the evidence base, not a sign that the thesis is safe.

There is a structural reason this matters for a body of work rather than a single piece: evidence is what lets a corpus compound. Anecdotes do not accumulate into knowledge — each is a self-contained story that contradicts the next one as easily as it confirms it. But measured facts compose; the deployment cited in one piece becomes a data point in another, the productivity finding in one article constrains the economic claim in a third, and over time the corpus behaves like an evidence base rather than a collection of opinions. This is the quiet payoff of the evidence-over-anecdote rule: it is not only that each article is more trustworthy, but that the articles can build on each other, because they are all anchored to the same kind of checkable reality rather than to incompatible intuitions. A library of anecdotes is a pile; a library of evidence is a structure.

Conviction is earned at the anchor

The apparent tension in this principle is that ENSI writes with great confidence — declarative, unhedged, axiomatic — while also insisting on evidence over plausibility. The resolution is that the confidence is licensed by the evidence, not by the absence of it. The unhedged sentence is permitted only over a claim that has been anchored to something measured; everywhere else, the method softens to the defensible form or cuts. This is the opposite of the confident pundit, whose certainty scales with their ignorance of the evidence. ENSI's certainty scales with the strength of the anchor beneath it, which is why a piece like the eGovernment opportunities survey can be both bold and correct: every bold sentence sits on a real deployment, and the boldness is just the deployment stated without flinching.

The wager of this principle is that plausibility and truth are not the same thing and frequently point in opposite directions — that the claims most likely to be wrong are the ones that sound most right, because they were selected for sounding right rather than for being true. The method's defense is to refuse plausibility as a credential and demand measurement instead: trials, data, deployments, numbers with origins, and the strongest available dissent. Anchor every claim to what was actually measured. What sounds right is the bait; what was measured is the catch. The whole discipline is learning to tell them apart, and to write only from the second.

Further reading

• AI-Driven eGovernment: The Opportunities
• Impact of Generative AI on Worker Productivity
• The Economics of Infinite Intelligence
• Meta Science: The Systemic Failures of the Scientific Machine
• Epistemology: From Discovery to Justification
• National AI Strategies: The Common Areas