klappy.dev

The StylebookConstraint

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 AlsoSibling Constraints

  • Anti-Cache Lying — 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 — Prevents ODD from becoming heavy, coercive, or self-justifying.

Canonical Tie-In

This constraint exists because:

"Nothing exceeded the threshold."

Markdown · RSS · built from klappy/klappy.dev@b750d7d