benchmark-to-latent-labels

Latent labels separate prompt-side structure from response-side behavior by classifying benchmark labels into direct targets, derived targets, validation signals, or nuisance variables. This process creates a machine-actionable specification for mechanistic interpretability and confound planning.

Building a labeling ontology from benchmark metadata requires classifying native labels into direct versus derived targets, then freezing a first-pass schema. This creates a structured inventory separating prompt-side structure from response-side behavior for downstream interpretability tasks.

Creating labeling functions for a mechanistic interpretability benchmark involves converting the native labels into a latent-label specification. You classify variables, build a benchmark-specific ontology, and generate corresponding functions alongside a confound audit gap list.

Yes, latent-label generation supports confound auditing by classifying benchmark labels into validation signals and nuisance variables. It produces a confound plan and gap list during ingestion to identify potential interpretability confounds.

You need to freeze a labeling ontology schema during benchmark ingestion when preparing for mechanistic interpretability. Freezing the first-pass schema stabilizes the latent-label inventory, ensuring downstream labeling functions and confound audits operate on consistent targets.