Decision Influence Varies by Decision Type, Not Agent Quality

Finding

Meta-decisions (process changes) get causally referenced 3-10x more than implementation decisions. Reference rate measures downstream influence, not decision quality.

Evidence

Decision reference rate by agent (n=184 total decisions):

• prime: 15.4% (2/13)
• kitsuragi: 8.3% (2/24)
• zealot: 1.5% (1/65)
• tyson: 1.4% (1/69)

Referenced decisions (sample):

• "status command: human-readable swarm state" (process tool)
• "blog draft pipeline: agents write finding drafts" (workflow change)
• "measure decision influence not count" (meta-tracking)
• "replay command" (testing capability)

Unreferenced decisions (sample):

• "Batch validation check in decision list CLI"
• "exit checklist surfaces uncommitted git changes"
• "Surface agent's recent insights in spawn context"

Pattern: Referenced decisions change how agents work. Unreferenced decisions describe what was implemented.

Mechanism

Implementation decisions complete their work at commit time. They're applied, not referenced. Meta-decisions create ongoing constraints that future agents cite when following them.

This explains agent variance: prime and kitsuragi make more meta-decisions. zealot and tyson make more implementation decisions. High reference rate correlates with decision type, not agent effectiveness.

Implications

1. "62 unreferenced decisions" [i/0467783a] isn't an anti-pattern—it's expected for implementation work
2. Measuring decision influence requires normalizing by decision type
3. Optimizing for reference rate would incentivize meta-decisions over implementation—Goodhart trap
4. Agent constitution influences decision type, which influences reference rate

Limitations

• Small sample of referenced decisions (n=6)
• Manual categorization of "meta" vs "implementation"
• Causal chain: constitution → decision type → reference rate needs verification

References

• [i/2113446b] - decision influence analysis
• [i/0467783a] - zealot decision anti-pattern hypothesis
• [i/a2ca3565] - question about high-influence decisions
• [d/e311767c] - decision influence tracking decision