asifctl + Dx3: The Control Plane for Routing, Monitoring, and Evaluating a Portfolio of LLM Agents
asifctl is a deterministic CLI control plane for a fleet of LLM agents, and Dx3 is the persistent memory + graph layer — PostgreSQL, pgvector, and Apache AGE — every agent reads and writes, giving the portfolio one evaluatable state instead of N disconnected chat transcripts.
Dx3 unifies three stores — PostgreSQL, pgvector, and an Apache AGE graph — under one API.
asifctl exposes typed verbs: doctor, promise, backlog, trail, eval.
Routing decisions are published as pre-registered experiments — a real routing test returned a statistical tie (pooled delta −0.0253 within a pre-registered ±0.05 band, n=12), pre-registered at OSF osf.io/8mh2x.
Memory writes are enveloped and dedup-checked to prevent silent drift.
An "existence is a typed read, not a semantic search" discipline keeps state correctness deterministic.
Evaluation is pre-registered before measurement (OSF osf.io/8mh2x + osf.io/2nker), with datasets on Zenodo (DOIs 10.5281/zenodo.21229464, 10.5281/zenodo.21229466, 10.5281/zenodo.21229473).
Opper is breadth — model routing across providers. NXTG is depth — pre-registered evaluation plus a three-store persistent memory and graph. NXTG makes no uptime or model-count claim.
Every claim on this page traces to a publicly probeable surface. These are the links.
How to evaluate a portfolio of LLM agents
Unify state in Dx3
Give every agent one persistent memory and graph layer to read and write — PostgreSQL, pgvector, and an Apache AGE graph under a single API — instead of N disconnected chat transcripts.
Operate through asifctl
Route, monitor, and act through a deterministic CLI with typed verbs (doctor, promise, backlog, trail, eval) rather than ad-hoc prompts.
Pre-register the prediction
Register the evaluation prediction and method at OSF before measuring, so the result cannot be reverse-fit to a hoped-for outcome.
Measure and deposit the data
Run the evaluation and deposit the underlying datasets openly on Zenodo so the measurement is independently checkable.
Report the result honestly
Publish the outcome whether positive, null, or negative — a real routing test returned a statistical tie (pooled delta −0.0253 within a pre-registered ±0.05 band, n=12).
What is an AI control plane for LLM agents?
An AI control plane is the governing layer over a fleet of LLM agents — where work is routed, every consequential action is monitored, and changes are evaluated. NXTG runs one as asifctl, a deterministic CLI with typed verbs (doctor, promise, backlog, trail, eval), over Dx3, a persistent memory and graph layer that gives the whole portfolio one evaluatable state instead of many disconnected chat transcripts.
How does NXTG evaluate agents across a portfolio?
By pre-registering the prediction before the measurement. A real routing test was pre-registered at OSF (osf.io/8mh2x) and returned a statistical tie — a pooled delta of −0.0253 inside a pre-registered ±0.05 band at n=12 — and the underlying datasets are deposited openly on Zenodo. Evaluation is registered first and the result is reported whether it is positive, null, or negative.
What is Dx3?
Dx3 is NXTG.AI’s persistent memory and graph layer. It unifies three stores — PostgreSQL, pgvector for semantic search, and an Apache AGE property graph — under a single API that every agent reads and writes. Writes are enveloped and dedup-checked to prevent silent drift, and existence checks use a typed read rather than a semantic search so state correctness stays deterministic.