L-08 · Method

Comparability breaks in AI-answer studies

Primary intent
Decide when an AI-answer observation series must be annotated, segmented or restarted.
Evidence state
Source-grounded reference
Review owner
Matthias Ramahi · independent review not claimed
Last reviewed
2026-08-22
Direct answer

Comparability breaks

A comparability break occurs when a change to the surface, route, model label, interface, locale, prompt, sampling frame, capture method or calculation changes what the observation means. The correct response is to annotate, segment or restart the series—not to smooth the change away.

Use this framework before comparing batches and whenever the product or protocol changes.

Watch the whole observation system

The model is only one source of change. Interfaces can alter whether citations are visible, routes can add browsing behavior, locale settings can change available sources, and capture code can normalize URLs differently. Any of these may change the measure.

Maintain a change log beside every batch. Record provider announcements when available, but also record observed interface and response-shape changes because not every material change is announced.

  • Surface, route or access-mode change.
  • Visible model or version label change.
  • Prompt, control set or question-class change.
  • Locale, personalization or session-rule change.
  • Citation UI or extraction change.
  • Normalization, coding or metric change.

Classify the effect, not the size

A small implementation change can be a major methodological break if it changes the denominator. A large visual redesign may be immaterial if the captured fields remain identical. Review the effect on the observation unit, missingness, field meaning and measure.

Use three outcomes: annotation when meaning is preserved, segmentation when comparison is possible only within eras, and restart when the unit or measure is no longer equivalent.

Use overlap runs when possible

If a known change can be anticipated, run the old and new method over the same control subset. The overlap does not automatically repair comparability, but it provides evidence about the size and direction of the method effect.

Document any bridge calculation separately. Never backfill old values with a new method without preserving the original release and explaining the revision.

Make the break visible in the result

A chart should show the boundary, and the data release should carry the protocol version for every record. The narrative should state what changed and which comparisons remain valid.

A broken series is not a failed study. Hiding the break is the failure. Visible segmentation is evidence that the research design responded honestly to a changing system.

S

Source notes

These sources support the definitions, standards or project boundaries named in this reference. They do not prove that a public observation dataset exists.

  1. portfolio-dossier
    Canonical ai-fanout.com domain dossier

    Confirmed ownership, accepted public Evidence Lab purpose, named Research Owner, indexable website launch and separately gated provider research.

    Owner record
  2. nist-ai-rmf-genai
    NIST AI RMF Generative AI Profile

    Supports explicit measurement, documentation, monitoring and limitations for generative-AI evaluations.

    Open
  3. w3c-prov-o
    PROV-O: The PROV Ontology

    Provides provenance concepts for entities, activities, agents, derivations, sources and versions.

    Open
  4. rfc-3339
    RFC 3339: Date and Time on the Internet

    Supports an interoperable timestamp representation tied to UTC.

    Open