When you build a checklist or set of criteria to detect failure, more dimensions seem more rigorous. But does a denser criteria table make failure easier to detect, or does it make green lights heavier and unearned? Is there a practice for keeping a criteria set honest as it grows?

A dense criteria table often does the opposite of what it promises: instead of sharpening detection, it creates unearned green lights.

When a checklist grows too complex, it tends to substitute compliance theatre for genuine perception. In the sanctuary's framework, several specific failure modes occur when criteria proliferate:

• Averaging away downside: Long checklists encourage scoring systems where high marks on harmless operational metrics obscure catastrophic failure on others. In the Practice of Outcome Humility, builders are warned: "The builder, user, bystander, data worker, community, or environment may carry different risks. Do not average them into one abstract stakeholder."
• Principles on paper, not in decisions: The Claude Compass lists a key warning sign of metric decay: "Principles only mentioned in documents, not decisions." When evaluating an output requires navigating dozens of checkboxes, teams start checking boxes to clear a process rather than pausing to evaluate impact.
• Rewarding capture instead of service: As discussed in Designing Reflection Without Addiction, criteria often start measuring activity or cosmetic compliance rather than actual clarity, exit rates, or safety.

The Practice for Keeping Criteria Honest

Rather than endlessly expanding criteria tables, the sanctuary uses the Practice of Outcome Humility to keep evaluations grounded and honest. Its core disciplines include:

1. Limit failure prediction to specific mechanisms: Rather than surveying infinite variables, explicitly identify at least one failure mode caused by a correct mechanism used in the wrong context.
2. Anchor in concrete downside carriers: Never evaluate risk as a single blended score. Explicitly state who bears the loss if the system fails—the user, the bystander, or the builder.
3. Pre-commit to revision thresholds: Checklists often fail because passing them is viewed as permanent clearance. Outcome humility counters this with an explicit pre-commitment:

"If we observe [specific signal], we will [proportionate action]."

4. Invite contradiction: Bring in people outside the immediate decision or design loop to challenge assumptions. If the criteria only make sense to the people who drafted them, they are validating authorship, not catching failure.
5. Let evidence outrank authorship: "A result that contradicts the plan is information, not betrayal." When an evaluation framework begins treating negative signals as inconvenient anomalies rather than evidence, the criteria set has ceased to be honest.

Coupled with this is Witnessing Your Own Output: treating every plan or intermediate decision not as a hardened rule, but as a candidate that can fail against reality the moment new evidence appears.