Evidence

Origin & evolution

The idea that decisions⚖️DecisionStrategyA recorded decision with context, rationale, and consequencesView reference → should rest on graded observation rather than authority comes from evidence-based medicine, formalised in the early 1990s, and the term migrated into management through Jeffrey Pfeffer and Robert Sutton's *Hard Facts, Dangerous Half-Truths, and Total Nonsense* (2006). Their argument was that organisations routinely act on conviction, copied practice, and casual benchmarking, and that treating belief as a thing to be evidenced is a competitive advantage.

In product practice the concept sharpened through experimentation. The lean startup loop made evidence the output of a test rather than a thing you go looking for to confirm a decision already made. The build-measure-learn cycle in Eric Ries's *The Lean Startup* (2011) treats each experiment🧫ExperimentValidationA test designed to validate a hypothesisView reference → as a generator of evidence about a leap-of-faith assumption🤔AssumptionStrategyA belief taken as true that underpins a strategyView reference →. Teresa Torres later pushed teams to weigh evidence by how it was gathered in *Continuous Discovery Habits* (2021), warning against confirmation-driven reading where a team counts the quotes that agree and ignores the ones that do not.

Annie Duke's *Thinking in Bets* (2018) frames this from the decision side: because outcomes🏆OutcomeStrategyA desired business or user outcomeView reference → are partly luck, the only honest unit of analysis is the quality of the belief at the moment the decision was made, and that quality is precisely what evidence is meant to track. Duke identifies a complementary failure mode to confirmation bias — "resulting," the tendency to let a good outcome retroactively validate weak evidence, or a bad outcome discredit strong reasoning. By that reading, recording evidence with an explicit weight and direction is a guard against resulting as much as against confirmation bias: the record fixes what the evidence actually said before the outcome was known.

Two refinements settled the practice. First, evidence has a direction: it can support, refute, or stay neutral on the claim it touches, and refuting evidence is often the most valuable. Second, evidence has weight, and that weight tracks rigour. A controlled experiment outranks an observation, which outranks a single quote, which outranks an expert's opinion.

How it works in practice

A team believes its trial-to-paid conversion stalls because pricing is unclear. They gather evidence against that hypothesis. An instrumented experiment shows users who view the pricing page convert at the same rate as those who do not: rigour quantitative, source experiment_run, direction refutes, weight strong. Eight interviews surface that users abandon at the point of inviting a teammate: rigour qualitative, source interview, direction neutral on pricing but supporting a new seat-invitation hypothesis. One loud support email blames price: rigour anecdotal, weight weak. Read together, the strong refuting evidence kills the pricing theory, and the qualitative cluster redirects the team toward collaboration. The anecdote is recorded, weighted low, and overruled.

Evidence vs. its neighbours

• Insight🔦InsightUser ResearchA synthesised finding from researchView reference →: evidence is a single observation with provenance; an insight is the interpreted pattern that several pieces of evidence add up to. Evidence is "six of eight users skipped the tutorial🧑‍🏫TutorialCustomer EducationA step-by-step tutorialView reference →." The insight is what that means and what to do. The graph routes evidence into interpretation through evidence_interpreted_as_learningEvidenceinterpreted asLearningcausal.
• Metric🔢MetricStrategyA unified metric that measures progress, health, or behaviour across the productView reference →: a metric is a standing measurement tracked over time; evidence is a discrete finding tied to a specific claim. A metric tells you the number moved; evidence tells you why a particular belief should change.
• Hypothesis: a hypothesis is the falsifiable claim; evidence is what bears on it. The relationship is directional, which is why evidence records supports, refutes, or neutral rather than simply accumulating in favour.

In the graph

In the Unified Product Graph, evidence🔎EvidenceValidationData supporting or refuting a hypothesis lives in the Discovery, Research & Validation region, in the validation domain, fourth in the hypothesis-anchored sequence. It is produced, never freestanding: inbound edges include hypothesis_has_evidenceHypothesishas evidenceEvidencehierarchy, experiment_run_yields_evidenceExperiment RunyieldsEvidencecausal, and experiment_produces_evidenceExperimentproducesEvidencecausal, so every piece traces to the thing that generated it. Its direction, evidence_rigor, and weight properties make a body of evidence sortable by stance and trustworthiness instead of by volume. Outbound, evidence_supports_opportunityEvidencesupportsOpportunitycross-domain and evidence_interpreted_as_learningEvidenceinterpreted asLearningcausal carry the finding upward into decisions. Evidence with no inbound provenance edge is, by construction, an unsourced claim the graph will not let you hide.