ai-fanout.comOpen the lab
Menu
Research library13 maintained references

Build the evidence before the trend.

Definitions, measurement contracts and field procedures for studying public AI answers without turning inference into fact.

Public maintained library · reviewed 22 Aug 2026

DefineObserveComparePublish
01

Reference index

Each route owns one user job. Filters change this index only; every reference remains available as a normal link.

  1. L-01ConceptQuery fan-outDefine query fan-out without claiming access to private generated queries.
  2. L-02ConceptObservable evidenceDistinguish direct public-output evidence from inference and unknown system behavior.
  3. L-03MeasurementSource diversityDefine source-diversity measures that remain interpretable across repeated AI-answer observations.
  4. L-04MeasurementAnswer stabilityDefine answer stability without reducing it to one opaque similarity score.
  5. L-05MeasurementCitation persistenceMeasure citation recurrence without equating persistence with source quality or ranking causation.
  6. L-06MethodSampling AI answersDesign a repeatable AI-answer sample that supports a defined research question.
  7. L-07MethodControl question setCreate a control set that detects change without drifting with every observation window.
  8. L-08MethodComparability breaksDecide when an AI-answer observation series must be annotated, segmented or restarted.
  9. L-09MethodMissing dataRecord missing AI-answer data without silently changing denominators or confusing absence with zero.
  10. L-10Data standardObservation schemaImplement a minimum viable record for repeatable public AI-answer observations.
  11. L-11Data standardReproducibility packagePackage methods, observations, transformations and limitations so another practitioner can audit the study.
  12. L-12Field guideAudit AI-answer sourcesRun a bounded, reviewable audit of visible AI-answer sources.
  13. L-13Field guideCompare answers over timeRun a transparent time comparison of public AI answers with explicit comparability and trend gates.
02

Use the right layer

  1. 01

    Start with a concept

    Fix the meaning and evidence boundary before choosing a metric.

  2. 02

    Choose a measurement

    Name the observation unit, denominator and interpretation limits.

  3. 03

    Adopt a method

    Freeze sampling, controls, missingness and comparability rules.

  4. 04

    Package the record

    Keep provenance, rights and corrections attached to the result.

This library explains how a defensible observation program should work. It is not evidence that collection has started or that a public trend exists.

Inspect methodology v0.1