Building reasoning systems that audit themselves.
Ascendanti is the in-house research lab of Atlas Firm. We build local-first autonomous reasoning systems — the kind that produce evidence of their own quality rather than asking the operator to take their word for it. The work sits at the seam where cybernetics, control theory, and academic methodology meet. Every system the lab ships is bounded, traceable, and self-auditing by construction; nothing else holds up at scale.
The lab does not chase frontier scale. It builds bounded artefacts that run on hardware the operator owns, that can be audited line by line, and that produce a signed trace of every decision they make. Cloud is for distribution, not cognition. The thesis is that the next decade of useful AI will not be won by larger models but by tighter loops — and that the architecture of those loops is the work.
A system that cannot show its work cannot be debugged, and a system that cannot be debugged cannot improve. — First premise
A multi-agent research-paper authoring system for political science.
Eight specialised agents — spec-author, source-curator, note-taker, lit-reviewer, debate-mapper, prose-author, methodologist, adversary — coordinated through a span-traced orchestrator. Every claim is grounded in a card. Every card cites a source. Every run produces a signed trace of what each stage did, decided, and emitted. A calibrator scores the manuscript against twenty-four hundred and one IS / APSR golden-set papers across ten axes; the audit office runs seventy methodology rules and eighty adversarial attacks before a paper is judged near-final.
Sovereign runs entirely on the operator's hardware — Qwen3-Next-80B-Thinking, Command-R, an MLX-backed model rotation through llama-swap. No cloud. No telemetry home. The architecture is a direct response to its predecessor's failure mode: a single-call pipeline asking one model to do eight people's jobs in one shot, and rejecting itself one hundred percent of the time.
Six tenets shape every system the lab builds. They are not slogans; each names a constraint that, if violated, returns the system to a state we have already learned does not work.
Reasoning runs on hardware the operator owns. The cloud is for distribution, not cognition.
Every agent has a declared scope — what it reads, what it writes, what it must not touch. Discretion lives in the model; authority lives in the contract.
Every output is a span event. A system that cannot show its work cannot be debugged, and a system that cannot be debugged cannot improve.
Quality is measured against an external reference set, not declared by the model that produced it.
Friction at the right boundaries, not none. The operator should always know which side of the seam they are on — biological cognition, or machine.
The work is never finished. Every run should be better than the last on axes we can name. Read the full ethos →
A local-first intelligence substrate. The predecessor.
Eighty-seven engines on SQLite. An event-sourced spine. A ten-general strategy engine on the quaternion sphere. A measurement protocol that could test seventeen of its own hypotheses and refute eight of them. PLATO taught the lab what a bounded reasoning kernel actually needs in order to be useful, and where the limits of human-life-as-dataset are. Sovereign inherits its discipline; the architecture is otherwise distinct.