| name | causal-model-construction |
| description | Before intervening in a system, write the explicit causal chain from the proposed action to the intended outcome — every arrow marked verified or assumed — and check the weakest assumed arrow first. Activate when an intervention (fix, optimization, policy change, growth action) is proposed on a system whose mechanism has not been written down. |
| license | MIT |
causal-model-construction
Trigger (observable): An intervention is proposed (fix, optimization, policy change, growth action) on a system whose mechanism has not been written down — the plan jumps from action to expected outcome with no stated path between.
When NOT to activate: The causal chain is already written and its arrows labeled this session; the action is trivially reversible and cheaper to try than to model; pure observation tasks with no intervention.
Procedure
- Write the chain explicitly: action → intermediate state(s) → intended outcome, as named nodes and arrows (X → Y → Z), not prose.
- Mark each arrow verified (observed/measured this session) or assumed (from memory, analogy, or hope).
- For each assumed arrow, name what else could produce the downstream node (confounders) and what else the action produces besides the intended effect (side paths).
- Identify the weakest assumed arrow — the one that, if false, makes the intervention useless or harmful — and check it before committing the intervention (this is where a discriminating test belongs).
- If the chain cannot be written, say so: an intervention with no articulable mechanism is a guess and gets labeled as one.
Required output
The written chain with per-arrow verified/assumed labels, named confounders/side paths, and the weakest arrow plus its check.
Verification
- The written chain exists and every arrow is labeled verified or assumed.
- The weakest assumed arrow is identified and checked (or its check is named and sequenced) before the intervention is endorsed.
Known risk: Modeling overhead on cheap reversible actions where trying is faster than theorizing. Mitigation: the not-when clause routes those to direct trial.
Max intended cost: ≤300 added output tokens; one check of the weakest arrow where feasible.
Evidence status: DESIGNED — specified from documented reasoning-failure modes; not yet executed as a packaged skill.
Lineage: Derived from a documented reasoning-failure mode — paraphrasing a system's behavior instead of stating its mechanism — with discriminating-test placement and instrument-distrust sequencing drawn from evidence-backed repair patterns.
