Interpretation
What models say
We extract structured claims from model outputs so that observations can be compared consistently rather than judged from summaries alone.
SemanticRisk observes how AI systems interpret public website content, compares those interpretations over time, supports cross-model comparison where multiple model observations are available, and separates meaningful semantic differences from simple wording variation.
Current paid measurement uses 60 neutral buyer questions for Comprehensive Review, 30 for Monitor, and 40 for Compare & Monitor. Questions cover 10 buyer-decision dimensions and draw on public evidence, market category and trust/risk context without forcing the target company into discovery answers.
Observation totals are calculated from questions × provider lanes × repetitions. Each cycle retains its question plan, methodology version and provider/model identity. Compare & Monitor uses the primary company's governed context for its comparison set.
Only compatible cycles support longitudinal change claims. A material methodology change starts a new baseline. The original six-question, 24-observation self-test remains historical evidence and is not directly comparable with the current paid matrix.
We extract structured claims from model outputs so that observations can be compared consistently rather than judged from summaries alone.
Where multiple model observations are available, claims may be equivalent, possibly equivalent, narrower, broader, contradictory or unrelated. Shared vocabulary by itself is not treated as proof of equivalence.
Repeated observations identify interpretation drift, including cases where model interpretation changes even when normalized source content appears unchanged.
A rewritten paragraph can preserve the same proposition, while one changed number, industry label, attribution or scope qualifier can alter the meaning. SemanticRisk considers the type and context of a difference, not only how many words changed.
Different wording appears to preserve the same central proposition.
The difference may affect attribution, specificity, scope, classification or a verifiable fact, but the available evidence does not support a firm material verdict.
The evidence supports a meaningful change in the proposition, its boundaries or its factual content.
SemanticRisk reports how AI systems interpret available public content. It does not independently certify every extracted statement as legally, commercially or factually complete. The distinction is central to the product.
Calibration results are dated and based on the reviewed sample available at that time. New reviewed examples may change thresholds, classifications and published performance statistics.
Current public methodology snapshot: 3 October 2026. Historical calibration results retain their original dates and evidence boundaries.