Embedded Risk Analytics builds on a single claim, that an agent stays coherent only while its capacity covers the load it carries.
The claim
C ≥ Hext + Bρ
We write capacity as at least the exogenous load plus the reflexive burden, the cost an agent pays to stay consistent with what it already committed. When the load crosses that boundary, the agent's coherence degrades. The boundary follows from the information the agent must carry, so it holds independent of the model, and the read and the repair both measure against it.
The foundation
The preprint sets out the information-theoretic model behind the program. It derives the feasibility inequality and shows how reflexive burden grows as a system nears its stability limit.
Papers
The program whitepaper, which sets the feasibility inequality that separates task difficulty from reflexive burden.
A position paper on Jev as a verifier for long-horizon agents. Claim readers built on Jev missed silently dropped commitments, and a deterministic committed-state check caught them.
A position paper on metering an agent's re-grounding capacity and triggering on trajectory state a dashboard cannot see.
Stateless swarms hold reflexive burden down, and they stop backing their conclusions under load, with a sharp failure boundary appearing once memory returns.
A managed agent held coherence at roughly 16,680 tokens per trajectory against 113,636 for a raw-window baseline, as the paper reports.
Experiments
On NVIDIA OpenShell's multi-agent notepad example, six Codex workers on gpt-4o-mini reported writes as done after a tool result that said still running, and some of those writes never landed. Right Rudder's read flagged each report as it was made, and its repair in front of each report sent the workers back to their writes, with reported figures lacking a landed source falling from thirty-five to three across fifteen workers in four runs.
On the Synthetic Hospital benchmark, a diagnosis agent working eight patients per session carried diagnoses from earlier patients into later ones, and a repair in front of the submission step removed them and lifted the benchmark score by 0.06 to 0.10 on GPT 5.6 Luna and DeepSeek V3.2.
Recovering a single committed fact cost 155 times a deterministic lookup on Amazon Bedrock and 85 times on Microsoft Azure, as measured in the study.
Four of five vendors' flagship endpoints reason by default, which breaks measurement once the token budget runs tight.
The read separated steps that contradict a prior commit from steps that are simply hard.
We add new writing here as it publishes.
