Skip to content
Validity typeConstruct
Pass conditionThe circuit is the smallest set of components satisfying sufficiency; removing any member degrades performance
Evidence familyCausal (per-component ablation)
Minimum reportingPer-component ablation result for each member; confirmation no member can be removed without behavioral degradation
Common failure modeCircuit grown by correlation or attribution score without a per-component pruning pass

Minimality requires no redundant members. Every component must be individually necessary given the other components. If removing c_i while leaving all other members intact has no effect on performance, c_i is redundant and should be removed.

Formally: for circuit C = {c_1, …, c_n}, minimality requires that for every c_i in C, ablating c_i while leaving all others intact produces a performance decrease.

The than_factor_specialization.py script tests this directly for the 5 “than” factors (f121, f219, f421, f546, f608): if all 5 activate similarly across comparative, rather-than, and quantity contexts, they are a non-minimal set — one factor would be sufficient. If they specialize by context (the script’s SPECIALIZED verdict), each factor is individually necessary for its context type — the set is minimal and each member has a distinct functional role.

  1. Take proposed circuit C.
  2. Ablate only c_i (zero or resample).
  3. Measure performance on target task.
  4. If not degraded relative to full-circuit ablation: remove c_i from C.
  5. Repeat until no member can be removed.

Report the set of members pruned and ablation results for each.

Multiple minimal circuits can exist (backup name-mover heads in IOI). This is a finding about architectural redundancy, not a failure of minimality. Report both minimal circuits and describe their relationship.

  • List every component in the final circuit.
  • Per-component ablation result (method, metric, delta) for each member.
  • Set of components pruned during minimality pass, if any.
  • If no minimality pass was run, flag as open criterion.

Discovery methods (ACDC, EAP, weight classifier) return components ranked by attribution score. Taking top-k without per-component ablation check risks including redundant members.