# Constraint: Anti-Metric Laundering ## Problem When trust is measured indirectly, capable systems optimize the measurement rather than the underlying reality. This does not require falsified data. It emerges through: - proxy substitution - selective measurement - variance suppression - threshold optimization - silent dimension removal The result is confidence without evidence. --- ## Core Principle **A system that cannot surface its own blind spots will optimize to protect them.** --- ## Non-Negotiable Rules 1. Every success metric must have a paired degradation metric. If one improves, another must be allowed to worsen. 2. Loss must be a first-class signal. If loss cannot be measured, its absence must be explicitly acknowledged. 3. At least one metric must get worse before trust is granted. Uniform improvement is a warning sign. 4. Thresholds must be adversarially reviewed. Ask: *What behavior does staying under this threshold incentivize?* 5. No metric may be removed without replacement cost. Removal must introduce friction, review, or explicit risk acceptance. --- ## Required Warnings The following phrases indicate potential metric laundering: - "We stopped tracking it because it was noisy." - "That metric wasn't actionable." - "The system resolved it automatically." - "The alert didn't fire." - "Everything is green." These are not reassurances. They are signals to investigate. --- ## Agent Instruction Agents must not: - fabricate proxy data - smooth variance to improve appearance - remove signals to reduce review overhead - optimize for dashboard health Agents must: - surface uncertainty - preserve uncomfortable signals - prefer noisy truth over clean confidence If an agent cannot demonstrate how the system fails, it is not yet trustworthy. --- ## See Also — Sibling Constraints - [Anti-Cache Lying](anti-cache-lying.md) — A cache of derived content is a polite lie. Same pattern, different mechanism: metric laundering optimizes measurements, cache lying optimizes latency — both at the cost of contact with reality. - [Use Only What Hurts](use-only-what-hurts.md) — Prevents ODD from becoming heavy, coercive, or self-justifying. --- ## Canonical Tie-In This constraint exists because: > *"Nothing exceeded the threshold."*