Game-Oriented Lagrangian Agent Governance — the research program behind quadratic voting, evolutionary calibration, and the act vs escalate gate used across Arkivist.
Bench: GOLAG
Overview
GOLAG (Game-Oriented Lagrangian Agent Governance) is the mechanism by which Arkivist achieves verifiable intelligence. Agents have finite confidence budgets; decisions optimize L = (Confidence × ContextMatch) / Risk.
Why it matters
Unchecked autonomy scales hallucination. Unchecked human review scales cost. GOLAG finds the pareto frontier: act when proof is strong, escalate when tension or risk is high — with an auditable trace either way.
Methodology
- Quadratic voting across 30+ domains (verification, federation, code, security, voice, and more)
- Replicator dynamics:
AC(gen+1) = AC(gen) + α(AC_best - AC_avg) - Expert agents (95%+ over 20+ decisions) earn budget bonuses; dying agents transfer wisdom
- ECE-style calibration measurement on every domain bench (Chess, Cyber Gym, Finance, SWE)
Results & next steps
GOLAG is not a slide — it ships in production pilots. Research papers and internal specs live alongside open benches so customers can reproduce calibration gains on their own VMs.
Arkivist Research
Updated February 1, 2026
