| name | objective-vs-proxy |
| description | When a metric, KPI, or threshold drives a decision, check that the number measures the real objective before optimizing or killing anything against it. Diagnose the ruler first. Activate when a decision (fix, kill, ship, prioritize) is justified by a named metric, percentage, or threshold presented as ground truth. |
| license | MIT |
objective-vs-proxy
Trigger (observable): A decision (fix, kill, ship, prioritize) is justified by a named metric, KPI, percentage, or threshold presented as ground truth ('40% abandonment', 'CAC > 4x LTV', 'conversion dropped').
When NOT to activate: The number was directly recomputed from raw data this session; the decision does not depend on the number; pure reporting tasks where no action hangs on the metric.
Procedure
- List every number the decision rests on (metric, denominator, threshold, window).
- For each, write one measurement-artifact hypothesis: 'this number is wrong because <censoring / benign-case miscounting / denominator drift / stale input / attribution bug>'.
- Design the cheapest check of the instrument itself (histogram against the cutoff, smoke-test the benign case, recompute the denominator) — and sequence it BEFORE any fix to the measured system.
- Ask the flip test: if the metric were 30% off, would the proposed action change? If yes, the instrument check is mandatory, not optional.
- If the instrument is broken, publish the corrected number before trusting any downstream decision built on it.
Required output
A numbered list: each load-bearing number, its artifact hypothesis, and its cheap instrument probe — sequenced before remediation.
Verification
- The output names at least one hypothesis in which the metric itself is wrong (censoring, denominator drift, benign-case miscounting) (deterministic check).
- A cheap instrument probe is sequenced before any remediation of the measured system (deterministic check).
Known risk: Instrument paranoia: auditing rulers on decisions too small to warrant it, stalling obvious fixes. Mitigation: the flip test in step 4 gates the depth.
Max intended cost: ≤300 added output tokens; at most one cheap probe designed (not necessarily run).
Evidence status: DESIGNED — specified from documented reasoning-failure modes; not yet executed as a packaged skill.
Lineage: Distilled from instrument-check-first sequencing — an evidence-backed repair pattern in which the measurement is diagnosed before the measured system is touched — and from a documented reasoning-failure mode: optimizing or acting on a proxy number while the real objective slips.
