Skip to content

An evidence family classifies a metric’s output by its source (where the signal comes from) and its access mode (whether the analyst observes or intervenes). The classification sits between metrics and criteria in the evaluation pipeline: metrics produce raw measurements, evidence families describe what type of epistemic content those measurements carry, and criteria then evaluate whether that content meets the bar for a validity claim.

Evidence divides along two axes. The source is the part of the system being measured: its persistent organization (weights), its runtime state (activations), its input-output relation (behavior), or its developmental history (training). The access mode is whether the analyst reads the system passively (observational) or perturbs it under a do-operator (interventional).

ObservationalInterventional
WeightsSVD, effective rank, OV/QK composition, MDLCircuit transplant, weight editing, weight knockout
ActivationsProbing, CKA, mutual information, PID, logit lensActivation patching, DAS-IIA, ablation, steering
BehaviorLogit diff, behavioral profiling, task accuracyPrompt perturbation, dose-response, KL under ablation
TrainingLoss curves, checkpoint comparison, phase transitionsAblate-then-retrain, fine-tuning, curriculum manipulation

Most published mechanistic interpretability work lives in one cell: Activations × Interventional. Activation patching, DAS-IIA, ablation, and steering are all interventional measurements of internal state. The framework does not privilege this cell — it notes that a claim supported by evidence from multiple cells is stronger than one supported by multiple methods from the same cell, because different cells have structurally different failure modes.

An earlier version of this framework used six evidence families (Causal, Structural, Representational, Behavioral, Information-theoretic, Measurement-theoretic). That taxonomy mixed source and method type: “Causal” spanned activations and behavior, “Information-theoretic” was a method applicable to any source, and “Measurement-theoretic” was a meta-level calibration applicable to all eight cells. The 4×2 matrix separates the axes cleanly.

The old families map approximately:

  • Structural → Weights (observational)
  • Causal → Activations (interventional) + Behavior (interventional)
  • Representational → Activations (observational) + Activations (interventional)
  • Behavioral → Behavior (both modes)
  • Information-theoretic → cross-cutting method applicable to any source
  • Measurement-theoretic → cross-cutting calibrations (bootstrap stability, seed variance, baseline separation) that apply to all 8 cells

How evidence families enable convergent validity

Section titled “How evidence families enable convergent validity”

The primary function of the family classification is to structure convergent validity assessments:

  1. How many cells support the claim? A claim supported by three cells is stronger than one supported by three metrics from the same cell.
  2. Which cells are represented? Weights × Observational combined with Activations × Interventional is powerful because their failure modes are complementary: runtime compensation cannot confound weight-space analysis, and weight-space structure that exists but is never activated is caught by interventional testing.
  3. Are there cells that should support the claim but do not? Absence of expected evidence from a cell is informative.

Evidence families are Layer 2 of the six-layer evaluation pipeline. After declaring the description mode (Layer 1), you identify which evidence families are relevant to the claim. This determines which metrics to run in Layer 3, which in turn produces the measurements scored against the 36 criteria in Layers 4–5.