Evidence and Confidence
Guide status: Foundation contract. Last reviewed 14 August 2026. This framework requires analytics, privacy and legal review where it is used in regulated or high-impact decisions.
Purpose
This guide explains how Liva 7 should describe evidence and confidence. The aim is to make a useful conclusion without hiding where it came from, how much uncertainty remains or what the design cannot prove.
Four evidence classes
Observed evidence is directly recorded for a defined source, scope and window: for example, an eligible item was available or a valid event was received. Observation can still be incomplete or affected by collection quality.
Derived evidence is calculated from observed inputs using a documented rule. A rate, grouping or deterministic relationship is derived. Its quality depends on the inputs and definition.
Inferred evidence estimates a likely state or relationship from patterns or a model. It may support prioritisation or personalisation, but must remain labelled as inference.
Experimental evidence is produced through a controlled comparison intended to support a causal decision. Its strength depends on allocation, exposure, integrity, duration and analysis.
What confidence means
Confidence describes the support for a particular conclusion under a stated method. It does not convert inference into observation or a comparison into proof. A high-confidence modelled relationship is still modelled; a low-volume observed count is still observed.
Where Liva 7 shows a confidence label or score, the method, inputs, freshness and material limitations should be available. Avoid implying precision that the underlying method cannot support.
Match the decision to the evidence
- Use observation to describe current state and identify questions.
- Use derived measures to compare consistently defined behaviour.
- Use inference to prioritise reversible work or decide what to test.
- Use experimentation when causal confidence materially affects the decision.
Increase the required evidence with impact, irreversibility, customer risk and cost. A reversible content ordering can tolerate more uncertainty than a material price, privacy or billing decision.
Preserve limitations
Record missing sources, small samples, selection effects, seasonality, unusual promotions, stale data and overlapping changes. A limitation should follow the evidence into the recommendation, decision receipt and later outcome review.
Safe use
When evidence classes conflict, do not average them into false certainty. Resolve the scope or source difference, or keep the conclusion unsettled. When evidence is insufficient, use Learning or Unknown, retain the safe default and choose a reversible next step.
You are finished when
- the evidence class, scope, source and window are clear;
- confidence is presented with its method and limitations;
- the claim is no stronger than the evidence; and
- the decision standard reflects impact and reversibility.