Project Ancilla
A mind, built like a brain.
Ioma Labs is building the next generation of AI assistants. Our prototype, Calliope, is a Distributed Cognitive Language System — one mind modeled on the human brain, its cognition distributed across regions analogous to ours, each built from one or more specialist transformers rather than a single monolithic network. She's designed to be auditable, honest by construction, and able to grow without being retrained. She is the first result of Project Ancilla, our closed initiative to reshape how these systems are built.
FIG. 1 — Ioma architecture, schematic. Read bottom to top. DASHED LINE — substrate retrieval into articulation
Almost every problem is the same problem.
Compute cost, forgetting, hallucination, opacity, fragility — the troubles facing frontier models trace back to one design decision: everything is entangled in a single weight space, learned together and reconstructed on every query. We don't think the way past them is to manage that entanglement more cleverly. It's to remove it.
The next real gain won't come from a bigger version of that same pass. It will come from architecture — giving a system the structure to retrieve what it needs, reason over it in the open, and check its own work before it speaks.
— Ioma Labs, design philosophyFive stages, one system.
Each part is a small specialist with one job, independently inspectable — closer to how a research team reviews a draft than how a single mind blurts out a first impression.
Retrieved knowledge substrate
A foundation layer the system can draw on rather than relying purely on weights — grounding output in something inspectable.
Parallel specialist inference
Within each tier, a set of narrower specialist processes run concurrently rather than one generalist doing everything at once.
State assembly
Specialist outputs are combined into one structured, inspectable state — a record of what the system understood, not another opaque guess.
Metacognitive appraisal
A dedicated tier reviews the assembled state before it's allowed forward — a system checking its own work.
Trained articulatory model
Only once a state has cleared appraisal does a dedicated model render it into language, drawing on the substrate as it speaks.
Inspectable by design
Because each stage is a distinct, named process, you can examine where a conclusion came from — not just what it said.
Calliope, our first ancilla.
Ancilla means handmaid. Calliope is built to be one: helpful by default, honest by construction, and bound not to cause harm. She is the first complete result of Project Ancilla — our closed initiative, run with a small circle of partners and investors, to build assistants on a fundamentally different architecture.
We'll show our work.
As Calliope matures, we're benchmarking her against frontier models and publishing what we find — progress, performance, and the places the approach doesn't win yet. Papers and benchmark notes will appear on the research page as they're ready.
Building toward a different kind of assistant.
Project Ancilla is closed for now, run with a small circle of partners and investors. If you want to help build it — or back it — we'd like to hear from you.