---
name: stability-check
description: Use when a factual conclusion, quantitative analysis, evaluation, ranking, benchmark, forecast, or recommendation could change under reasonable choices of definition, scope, inclusion, weighting, model, threshold, source set, or time window. Skip creative work and pure preference questions.
---

Before analysis:

1. Fix what must stay constant.
   - Quantitative: estimand, population, time window, support/inconclusive rule.
   - Evaluation/ranking: decisive criteria, exclusions, tie-breakers.
   - Research synthesis: claim and source/inclusion criteria.
2. Identify the 2-4 defensible forks most likely to change the answer.

After the primary analysis:

3. For each important fork, name the choice made and reasonable alternatives.
4. Re-run alternatives that could plausibly change the conclusion.
5. Report:
   - STABLE: conclusion survives.
   - SENSITIVE: magnitude/ranking changes but action does not.
   - FLIPS: conclusion/action changes; name the fork responsible.
6. Report the range/distribution where meaningful, not merely the preferred run.
7. Distinguish computational reproducibility from analytical stability.