skills.rest
concordance-co
concordance-coOFFICIAL
Explore all skills by this author →
GitHub

mechanistic-interventions

Define intervention plans with target behavior, success criteria, and paired controls.

4|1|Updated May 20, 2026
Search Tags:#interpretation#interventions#experimental-design#causal-analysis#activation-patching#read-vs-write#site-selection
One-click install
npx skills add https://github.com/concordance-co/xenon --skill mechanistic-interventions
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: mechanistic-interventions
Source: https://github.com/concordance-co/xenon/tree/main/.agents/skills/mechanistic-interventions
Command: npx skills add https://github.com/concordance-co/xenon --skill mechanistic-interventions

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Move from benchmark-driven analysis into causal or mechanism-oriented interventions. Covers activation patching, interchange, control design, read-vs-write distinctions, and intervention-specific success criteria, while pointing to future attention and routing follow-up work.

Core Features & Use Cases

  • Intervention framing: define exact behavior to change, success criteria, malformed-output criteria, and intended direction of change.
  • Site choice from the computation story: select plausible sites based on localization cues and timing, not ease of patching.
  • Paired design and controls: prefer matched donor-target pairs, single-layer first tests, same-label controls, and purposeful control strategies.
  • Interpretation discipline: track intended-direction flips, reverse-direction flips, malformed outputs, and same-label instability; separate causal evidence from broad destabilization.
  • Follow-on mechanism work: point to likely next steps such as attention follow-up, routing/MoE follow-up, narrower span decomposition, and read-vs-write comparisons.

Quick Start

Draft an intervention plan detailing the target behavior, success criteria, plausible intervention sites, and a paired-control design.

Frequently Asked Questions about mechanistic-interventions

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I design a causal intervention plan for activation patching?▼

A causal intervention plan specifies the target behavior, success criteria, and plausible sites based on localization and timing. It outlines a paired-donor target design with controls to support causal inference during activation patching.

What is the read-vs-write distinction in mechanistic interpretability?▼

The read-vs-write distinction separates observing activation patterns from actively modifying them. Tracking this difference ensures interpretation frameworks correctly identify causal evidence rather than broad destabilization from the intervention.

How do I choose intervention sites for causal analysis?▼

Choose intervention sites based on localization cues and timing within the computation story, not ease of patching. This ensures targeted causal analysis accurately isolates the mechanism responsible for the behavior.

What controls do I need for a paired-donor experimental design?▼

A paired-donor experimental design requires matched donor-target pairs, single-layer first tests, and same-label controls. These purposeful control strategies help separate true causal evidence from unintended destabilization.

How do I interpret activation patching results without confusing causal effects with destabilization?▼

Interpret activation patching results by tracking intended-direction flips, reverse-direction flips, and malformed outputs. This discipline separates genuine causal evidence from same-label instability or broad destabilization caused by the intervention.

What follow-up mechanism work should I plan after initial causal interventions?▼

Follow-up mechanism work includes attention follow-up, routing or MoE follow-up, narrower span decomposition, and read-vs-write comparisons. These steps refine the causal analysis and pinpoint next mechanism-focused targets.

Related Skills

Explore More Data & Analytics Skills→
gagan114662gagan114662

pyvene-interventions

Perform causal interventions on PyTorch models using a declarative dict-based framework.

Community
Advanced
MGPowerlyticsMGPowerlytics

strategic-thinking

Enforce first-principles reasoning with decomposition, reconstruction, and verification.

Community
Basic
ccashwellccashwell

single-subject-design

Identify and implement single-subject experimental designs for behavior-analytic evaluation.

Community
Advanced
nahisahonahisaho
3

scientific-causal-inference

Estimate causal effects from observational data using PSM, IPW, DID, and RDD.

Community
Advanced