A knowledge graph that grades its own evidence.

Four hundred and twenty sources. Every one DOI-verified. Built to separate what behavioral science actually shows from what design decks have been repeating for a decade.

“Nudge” became a word before it became a discipline.

Behavioral-science vocabulary spreads through design faster than its evidence does. Named “laws” get cited in decks, style guides, and design documentation, and almost nobody checks whether the research behind them actually holds up.

This graph exists to answer that question, for every one of them, on the record — and to keep the answer current as new evidence shows up.

Three tiers. One direction only.

The arrow only runs one way. A real-world example can illustrate a bias. It can never inflate the rating above it. Popularity in practice and rigor in research stay visibly, permanently separate.

Not a prototype.

Hosted on Neo4j Aura, always-on and reachable from anywhere, not a local file on one laptop.

A map, not a spreadsheet.

The rule is simple. Holding it is the hard part.

Every source needs a real, checkable DOI. Every finding gets one of four grades — replicated, disputed, undetermined, or not applicable — never just “cited.” Every write to the live database runs through a validation pass before it touches anything. Fields are closed-enum, not free text, so “sort of backed, I guess” was never an option to begin with.

That discipline is the entire point. Picking quotes that sound authoritative is easy. Holding a consistent bar for what actually counts as evidence, and writing it down every single time, is the part that's usually skipped.

Scarcity Principle, start to finish.

Every link in that chain is a real, independently checkable citation, not a paraphrase of “research shows.”

When the evidence isn’t there, the graph says so.

Running the same audit against fifteen widely-cited “laws” — Murphy’s Law, Occam’s Razor, Jakob’s Law, and the Law of Triviality among them — found no rigorous backing behind any of them. That doesn’t mean the underlying observations are wrong. It means the evidence usually attached to them isn’t there when you actually go look for it.

A reliable answer nobody can find isn’t reliable.

Semantic search over the findings, tested against a labeled set of real questions rather than eyeballed. The production configuration beats two reasonable alternatives by a real margin.

One result changed an approach already shipped: a re-ranking step that clearly improves search over the findings makes search over the named biases worse. A technique that works for one part of a system doesn’t automatically work for another part just because it’s the same system.

Ask it a question.

Built to grow. Built to be checked.

Right now, this is mine: I query it, I maintain it, I’m the only one who has ever written to it. The current stretch of work is depth, not breadth: extending source coverage on the corpus’s most-cited, most foundational names, so the concepts practitioners lean on hardest rest on more than a single study.

Eventually, I want other designers to explore this themselves rather than take my word for what’s in it. That’s deliberately not built yet. When it is, I want it testable, not just readable — somewhere anyone working in this field can push on a claim and see exactly what backs it.

If you’ve ever cited a “law” you’d never checked, you’re not alone.

That's exactly why I built this. I'm always glad to talk through what's behind a specific claim, walk through the graph directly, or hear about a source I'm missing.