Skip to content

C — How to Pick a Mode Tag: Decision Tree

Section titled “C — How to Pick a Mode Tag: Decision Tree”

Use this decision tree before writing any verdict.


Start: What is the claim about?
├─► "The model predicts / performs / scores on task X."
│ └─► [functional]
├─► "Component X encodes / represents information about variable Y."
│ └─► M3 (baseline-separated IIA) established?
│ ├─► Yes → [representational]
│ └─► No → [functional]
├─► "Component X causally implements / is necessary for computation Y."
│ └─► I1 (necessity) established?
│ ├─► No → use [representational] or [functional]
│ └─► Yes → I2 (sufficiency) established?
│ ├─► No → [causal-mechanistic] Causally suggestive
│ └─► Yes → [causal-mechanistic] Mechanistically supported or above
├─► "The weights of component X implement computation Y."
│ └─► C2 (structural plausibility) established?
│ ├─► No → use [representational] or [causal-mechanistic]
│ └─► Yes → I1 causal support?
│ ├─► No → [structural-mechanistic] Proposed
│ └─► Yes → [structural-mechanistic] Causally suggestive or above
└─► "The mechanism generalizes across [model / task / prompts]."
└─► E5 (robustness) established?
├─► No → use non-transportable tag with scope restriction
└─► Yes → [transportable]

What researcher writesEvidence supportsCorrect tag
”L8.MLP implements SVA” (no circuit-only pass)I1 only[causal-mechanistic] Causally suggestive
”Circuit generalizes to other models” (no cross-arch test)Neither E5 nor E6[causal-mechanistic] Triangulated with scope restriction
”Factor represents subject-verb number” (IIA=0.48, no baseline)M3 partial[representational] Causally suggestive (pending baselines)

Mode tag (description level) and verdict tier (strength of evidence) are independent.

  • [representational] Validated — thoroughly validated representational claim
  • [causal-mechanistic] Causally suggestive — causal claim with only necessity established
  • [structural-mechanistic] Triangulated — structural claim corroborated by multiple instruments

Always state both.