Skip to content

Frameworks

Interpretive variability assessment

Framework for qualifying output variation across contexts without collapsing it into ranking, hallucination, or visibility folklore.

Also inFrançais
SectionFrameworks
Classificationnormative
Date2026-07-04

Interpretive variability assessment qualifies how and why generated interpretations differ across controlled contexts.

The framework exists to prevent over-reading isolated tests. It separates normal wording variation from variation that affects meaning, source authority, category, proof, or recommendation.

Minimum observation record

Each observation must preserve:

  • target entity, source, claim, product, or category;
  • canonical source or baseline, when available;
  • query and language;
  • system or surface observed;
  • web-search or retrieval status, when visible;
  • approximate context such as region, session mode, or device class, when relevant;
  • timestamp;
  • cited or activated sources, when visible;
  • raw output;
  • qualification decision and reason.

Assessment axes

IVA-1: retrieval variation

Cited or activated sources differ across contexts.

IVA-2: selection variation

Comparable sources exist, but different sources dominate the answer.

IVA-3: representation variation

The role, category, limits, proof, or differentiators of the target change.

IVA-4: recommendation variation

The target moves between absent, mentioned, recommended, first recommended, or displaced by a competitor.

IVA-5: context variation

Region, language, session state, tool mode, or time window changes the answer substance.

IVA-6: delivery-layer variation or fixation

The observed answer may be fresh, cached, routed, approved, refined, or assembled from an older retrieval state. Identical outputs can therefore indicate either stability or delivery-layer fixation.

Qualification states

  • Stable: wording differs, meaning holds.
  • Fragile: minor displacement appears but canon remains recoverable.
  • Unstable: category, proof, or recommendation changes materially.
  • Drifted: output contradicts canon or baseline.
  • Inconclusive: evidence is insufficient or context is uncontrolled.

Evidence rule

A single output cannot establish stability. A repeated pattern can support a stability, fragility, or drift finding only if the test conditions and raw outputs remain reconstructible.

Relationship to interpretive observability

Interpretive variability assessment is a component of interpretive observability. Observability provides the traces, metrics, and evidence needed to decide whether variability is noise, an early warning, or a governed drift condition.

Native vs delivered object

A valid assessment should state whether the observed object is a native model interrogation or an application-delivered answer. If the path is unknown, claims must be downgraded to delivered-surface observation.