assumption-audit

Purpose

The most dangerous claims in a PM's research corpus are the ones that feel obvious. A team sees an entity in the graph, treats it as fact, and plans against it — never noticing it came from a single ticket, or from nowhere at all. This skill makes that invisible risk visible.

It returns every Assumption node (zero sources) and every node/edge with only one source of provenance, ranked by how much the team would be betting on it. Each item is tied to the decision it would affect, so the PM can decide which to validate before committing.

It does NOT resolve the assumptions. It does not say which are likely true or false. It does not propose tests, mock evidence, or assign probabilities. Surfacing the risk is the deliverable; deciding what to do with it is the PM's call.

Procedure

1. Read graph/graph.json. If it does not exist, tell the PM to run graphify first and stop.
2. Read meta.health. Cold-start branch: if health === "thin", state at the top that on a thin corpus most claims will be single-source — single-source is the default state, not an anomaly. The triage is still useful, but the bar for "this needs validating" resets accordingly.
3. Collect items to audit: a. Every Assumption node (zero provenance) — these are first-class risks by definition. b. Every node whose provenance array has length 1 — single-source claims. c. Every edge whose provenance array has length 1, if it contributes meaningfully to the graph's structure (e.g. contradicts edges with a single source, or blocks edges between major nodes).
4. For each item, infer the decision at risk — what product or strategy call would change if this claim turned out to be wrong? Examples:
   • "Mid-market wants SSO before invites" — affects scope of the onboarding admin walkthrough.
   • "Jira import is the cause of churn, not the trigger" — affects whether GA-ing the native Jira import alone will move conversion. If you cannot identify a specific decision at risk, the item is probably not load-bearing — but still list it as "low".
5. Rank items by risk. Three levels:
   • High — a major decision (scope of a quarter, a public positioning claim, a one-way-door commitment) depends on this being true.
   • Medium — affects prioritization or framing but is recoverable.
   • Low — interesting but not load-bearing.
6. For Assumption nodes specifically, copy the risk field if present, or write one in if missing. The combination of "no evidence" + "the decision at risk if wrong" is the entire point.
7. Emit the output contract.