RESEARCH
Papers
The Raising Agents research program publishes two papers and two long-form companion reports. Paper 1 is the empirical lemma: preregistered, multi-study, 3,797 runs, four domains. Paper 2 is the manifesto layer: a position whitepaper folding five formal traditions and five governance traditions into a single reliability frame for stochastic agents. The two hyperresearch reports (R1 literature review, R2 industry-discourse audit) are the citation graph behind the whitepaper. All artifacts are open.
-
Paper 1
Output-Passing Behavioral Regressions in AI Agents
Detection, Baselines, and Repair with Agent Behavior Contracts
91.4% of behavioral regressions produce passing outputs. Behavior contracts detect them at F1 = 0.982 against best-baseline F1 = 0.400. Five studies executed; two preregistered for replication. The first empirical lemma in the program.
-
Paper 2
Delegation-Grade Agents
Why Determinism Is Not Enough for Agent Reliability
Five formal traditions (PCTL, e-processes, robust MDPs, AI reliability science, ABC) and five governance traditions (Hawley, Bhardwaj, Kolt, Jarrahi–Ritala, Hadfield–Menell) converge on the same answer without coordinating across families: probability-bounded behavior under attributable commitments. Determinism is the degenerate corner of every one of them. The Drift Bounds Theorem quantifies the cost: thirty-fold compliance gap over 100 agentic steps.
Companion research reports
The two hyperresearch reports below were fired before Paper 2 was drafted, to avoid strawman positioning. R1 surveys six academic literatures; R2 audits eleven vendors. Together they supply the citation graph behind every load-bearing claim in the whitepaper.
- R1 — Determinism, Delegation, and Trust. A layered synthesis across six literatures: software testing under non-determinism, stochastic systems reliability, AI/agent evaluation, human-agent trust calibration, organizational delegation theory, philosophy of trust. ~50 cited primary sources. Verdict: determinism is the wrong top-level target; five traditions converge on probability-bounded specification.
- R2 — AI Vendor Non-Determinism / Evals Framing. How Anthropic, OpenAI, Google Vertex, LangChain, Braintrust, Langfuse, Arize Phoenix, DeepEval, Weave, AgentOps, and LangSmith frame non-determinism, reliability, and production readiness in 2025–2026. Verbatim quotes from primary product documentation, engineering posts, and cookbooks.
For teams
Use the papers as evidence, not as a consulting brochure.
The public papers define the failure class and the reliability frame. If your team needs to apply this to production agents, that work routes through Zartis. Public artifacts stay public; client traces, internal policies, and implementation details stay private.