causal-audit
Skill8 starsChanged 57 days agoMITView on GitHub
Correlation walked in; make it prove causation. Use when someone claims "X causes Y" from observational data, asks "did the change actually cause the improvement?", "is this confounded?", "what would have happened if we hadn't shipped it?", plans an A/B test, or investigates why an incident/outbreak spread the way it did.
In this repository16 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- boundary-designSkill
- decision-bias-checkSkill
- estimationSkill
- evidence-synthesisSkill
- experience-and-transmissionSkill
- failure-forensicsSkill
- formal-correctnessSkill
- measurement-disciplineSkill
- narrative-sensemakingSkill
- normative-designSkill
- problem-reframingSkill
- representation-and-possibilitySkill
- structure-discoverySkill
- systems-leverageSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 42
If no shape above matches, use a built-in agent with the closest GENIUS: pattern from rules/agent-routes.json instead.
- Distinctive: line 48
No falsifiable version of the claim can be stated.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Causal Audit
Problem shape: a causal claim is being made, tested, or acted on, and the evidence is correlational, confounded, unreplicated, or resisted by the organization even when clean. The move: make the causal structure explicit, then intervene or design the comparison that could refute it.
Relevant geniuses
| Agent | Use when |
| pearl | "X causes Y" claimed from correlation; counterfactual questions; controlled-for variables may be colliders — draw the DAG first |
| fisher | an experiment is being designed: randomize, block, replicate, factorial over one-at-a-time; "run it and see" needs a design document |
| peirce | surprising observation needs candidate explanations; several hypotheses and a limited budget — test cheapest-to-refute first |
| semmelweis | matched groups with wildly different outcomes; the evidence is clear but the institution resists it |
| snow | something is spreading (bug, outage class, behavior) — outbreak investigation, case definitions, Hill's criteria |
| mill | only a handful of cases exist — methods of agreement/difference, necessary vs sufficient conditions |
| feinstein | diagnosis under uncertainty: differential ranked by likelihood ratios, treatment thresholds before acting |
Invocation
- Pick the best-fit agent above. If two or more fit, run
tools/genius-invoker.sh route "<problem>"and take the top ranked match. - Load it:
tools/genius-invoker.sh invoke <agent> "<problem>", then readagents/genius/<agent>.mdin full. - Apply the agent's
<workflow>step by step and answer in its<output-format>. State explicitly which rung of the ladder (association / intervention / counterfactual) the final claim sits on. - Typical chain: pearl classifies the claim → fisher designs the intervention
→ semmelweis plans the communication. Run via
tools/genius-invoker.sh compose pearl fisher -- "<problem>". - If no shape above matches, use a built-in agent with the closest
GENIUS:pattern fromrules/agent-routes.jsoninstead.
Refuse when
- The requester wants a causal conclusion from data that can only support association — name the missing intervention instead.
- No falsifiable version of the claim can be stated.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.