causal-explanation

Causal Explanation

Answer the selected causal question with evidence calibrated to the claim. Keep observed facts, source-backed inference, competing explanations, and unknowns distinct. Do not turn an explanation request into implementation work.

Route The Request

• Bind the question to the exact behavior, decision, regression, threshold, or tradeoff. Infer the referent from current context only when the interpretation is safe; otherwise ask one targeted question.
• Use jcode for a symbol lookup, file map, caller trace, or adjacent source context. Use the repo's QA intake or attribution lane for reproducing, diagnosing, or attributing an active failure, including a request to find an unknown root cause. Use this skill after the cause or regression finding is established, or when the selected job is to explain existing evidence.
• Use the applicable architecture or performance skill when the requested outcome is a design review or benchmark judgment rather than an explanation.

Build A Proportionate Evidence Record

1. Inspect the current code, runtime state, configuration, or artifact that establishes what exists now.
2. Consult only sources likely to resolve the causal question. Use jcode for source structure and callers, jdocs for authored documentation or reference trees, direct QMD retrieval plus get or multi_get for exact prior wording or history, and jdata only when tabular evidence is material. Consult a relevant issue or review connector only when current evidence points there. Do not require an all-source sweep or enumerate and sweep connectors.
3. Record each source actually consulted. Mark relevant searches that returned nothing, unavailable sources that leave a material gap, and deliberately omitted categories whose evidence could not affect the answer.
4. For a regression, compare the closest defensible known-good state with the current state and inspect the exact intervening changes. Timing is a hypothesis, not proof of cause.