Research
How the infrastructure works
The ideas behind UluOps — six core innovations that make AI judgment observable, persistent, and self-improving.
Recursive Appreciation
Agents audit agents. A prompt engineering agent can audit its own definition, find real issues, and improve. Improvements to meta-layer agents cascade through every domain-layer agent they evaluate.
Cognitive Parallax
The same artifact examined through multiple epistemological traditions reveals structural properties invisible from any single perspective. Aristotelian, Popperian, Humean, and Confucian lenses produce triangulation signal no individual framework can generate.
Composition Lift
Workflow compositions of multiple agents produce synthesis scores above the average of their component scores. Compositions are first-class evaluation units with their own emergent epistemic behaviors.
Cross-Model Triangulation
A finding's fingerprint is built from the agent, the file, the category and the title, never the model. The same agent definition run on different models gives the same identity when it describes the finding the same way, so agreement across models carries more evidential weight than recurrence on one. The limit: a finding the agent re-describes looks new.
Decision History
Every change to a finding's status keeps the prior state, and the reason, in its history. A decision can be revisited without the earlier one disappearing, which is what lets the record be measured over time.
Domain Portability
The infrastructure remains constant; only the definitions change. The same agents, workflows, and measurement apparatus work across software, security, prompts, ML pipelines, and beyond.
Evidence
What the data shows
257
Agent definitions
109
Cognitive lens agents
6 of 17
Audited agents converged
5
Domains in the registry
Six of 17 audited agents met a fixed convergence criterion, each within 2–4 audit runs (median 3). Three were re-audited without meeting it, and eight have had a single audit. Meeting it once is not permanent: one agent that converged failed two later audits.
Published agents span five subject domains: software, security, legal, scientific and compliance, alongside general-purpose and cognitive-lens agents.
Convergence: a prompt-audit run in which prompt-engineer scores at least 85 and no auditor reports a critical finding. Counted over the 17 agents audited in their own tracker project, as of 2026-10-22. Registry numbers: published, public definitions as of 2026-10-23.
Open Questions
What we're still learning
Convergence theory
Under what conditions does recursive self-evaluation converge, and at what structural ceiling? Is the limit Godelian, or artifact-specific?
Meta-learning
Can the system learn which agents to create? What are the optimal structures for recursive improvement and automated domain adaptation?
Cross-paradigm synergies
Which philosophical and analytical frameworks compose productively? Are there emergent properties at certain composition depths?
See the infrastructure in action
Explore the documentation, browse the agent catalog, or start building.