The Composition Divergence Index measures mismatch between organic resolution and generative composition.
For a query q and entity e, the protocol computes:
e.e.The body defines CDI as:
CDI(q,e) = organic resolution strength − composed admission strength
A positive CDI indicates that the organic layer resolves the target more strongly than the generated answer admits it.
The default procedure is:
1. identify the target through name plus canonical identifiers; 2. record platform, interface, time, geography, and account state; 3. capture the top ten organic results; 4. classify each result as target, other entity, or ambiguous; 5. capture the full composed answer; 6. classify each substantive claim as about target, other entity, or neither/both; 7. calculate both strengths and their difference; 8. deposit the evidence, rules, metadata, and result.
The comparison substrate must be declared. Organic results are the default, but DOI sources, ORCID, Wikipedia disambiguation, or other canonical surfaces may be used for a different measurement.
The paper-reported worked example concerns the query “lee sharks” on Google AI Mode in May 2026 and reports a movement from maximal to partial divergence after a silent state change. These are dated case-study results supplied by the source work.
Failure modes include classification disagreement, ambiguous comparison substrates, mutable composition state, and small generated-answer samples.
CDI is designed for longitudinal use through the Drowning Test and as a diagnostic companion to SPXI and SIM.