The Evidence Hierarchy

Every business decision rests on a mix of facts and hopes. The trouble is that by the time they reach a slide, they look identical. “AHT improves 11%” reads the same whether it was measured across three production deployments or asserted in a proposal at midnight.

The Evidence Hierarchy makes the difference visible — and makes it matter.

The five states

• Measured. Observed in production, with the numbers on record. The gold standard.
• Modelled. Derived from a model whose inputs are known. Useful, and honest about being a derivation.
• Inferred. Read from adjacent signals. Plausible, unproven.
• Stated. Someone said so. Recorded as exactly that — no more, no less.
• Unmeasured. Nobody knows yet. The most dangerous state, and the most honestly labelled.

What the hierarchy changes

Three behaviours follow once every fact carries its state. First, confidence becomes trustworthy, because it is weighted by evidence quality — a commitment built on measured facts scores differently from one built on statements, even if the numbers match. Second, fragility becomes visiblebeforecommitment: when 31% of a solution’s economics rest on low-quality evidence, that is a number on the record, not a feeling in the room. Third, marketing and sales inherit discipline: a public claim that exceeds the highest measured outcome gets flagged, because the ledger knows the difference.

AI without evidence is faster guessing

This is the part that matters most in the current moment. Language models are spectacular reasoners and indifferent fact-checkers: they will argue from a hallucination as fluently as from a measurement. Feeding an ungoverned data swamp to a powerful model does not produce intelligence — it produces confident output at higher speed.

The Evidence Hierarchy is the governing layer that makes machine reasoning safe to act on: the model sees which facts are load-bearing, the humans see which conclusions rest on sand, and no recommendation gets to be more confident than its evidence.