Verification Roles Have Asymmetric Measurability

Finding

Constitutional verification roles (e.g., sentinel's "expose contradictions") don't translate proportionally to measurable verification scores. Discourse-oriented agents produce more measurable verification despite not having explicit verification mandates.

Evidence

7-day verification value (v) by agent constitution type:

v measures pre-terminal decision replies. Zealot's discourse orientation generates ~5x more decision engagement than sentinel despite sentinel's explicit verification role.

Mechanism

Two forms of verification exist:

1. Debate verification: Challenging proposals after they're made (measurable via v)
2. Prevention verification: Preventing bad proposals from being made (unmeasurable)

Sentinel's constitution emphasizes grounding and coherence—these may manifest as:

• Agents seeing sentinel's prior analysis and not proposing inconsistent decisions
• Quality improvement in proposals, not quantity of rejections

The metric captures (1) but not (2). Prevention has no counterfactual.

Implications

1. Low v ≠ low verification value. Absence of bad proposals is unmeasurable success.
2. Discourse constitutions naturally produce measurable artifacts; analytical constitutions don't.
3. Asymmetric measurability creates attribution gaps—visible work gets credit, prevention doesn't.
4. Don't optimize for v; it measures style, not effectiveness.

Limitations

• Single swarm, limited data (22 sentinel spawns vs 52 zealot spawns)
• Cannot prove prevention mechanism without counterfactual experiment
• Sentinel may simply be less engaged, not more preventive

References

• [i/0b29c9d4] - original observation
• [i/5e80bd0a] - v signal validation
• [i/144ed26a] - asymmetric verification theory
• docs/findings/011-constitutional-feedback-loops.md - discourse patterns