Calliope
Calliope, and Project Ancilla.
Calliope is our first complete ancilla — a Distributed Cognitive Language System we're bringing to maturity and benchmarking against frontier systems. She is the first result of Project Ancilla, our closed initiative to build assistants on a different architecture.
What Calliope is.
Calliope is modeled on the human brain. Where a conventional model packs the world, language, and reasoning into one weight space, Calliope distributes cognition across regions analogous to the brain's — each a specialist, some built from several transformers rather than one. They perceive, reason, and remember separately, then converge into a single voice.
Knowledge outside the weights
Calliope retrieves what she needs when she needs it, rather than relying on what's baked into her parameters — so claims trace to a source, and what she knows can grow without retraining who she is.
She reviews her own work
A dedicated layer appraises her reasoning before it becomes language — a structural checkpoint, not a prompt trick.
Auditable by design
Because she's built from distinct specialists, the work at each boundary can be examined — how a conclusion was reached, not just the final output.
What that changes.
Because her knowledge, reasoning, and language are separate rather than entangled, several things that are hard for a single-pass model come almost for free.
She grows without forgetting
New knowledge is added outside her weights, so learning something new never overwrites something old. She improves without being retrained from scratch.
She can say “I don't know”
Whether she knows something is computed apart from what the answer is — so honest uncertainty is a real signal, not a guess dressed up as confidence.
Most of the time, she's cheap
The expensive reasoning step is recruited only when a question needs it. Routine exchanges never pay for deliberation they don't use.
Where it stands today.
Calliope is in closed testing. We're running her against frontier single-pass models to understand where this approach genuinely helps, and where it still falls short.
-
Phase 1 — Complete
Architecture design & tier specification
The architecture defined end to end — the specialists, the retrieval seam, and the appraisal step specified.
-
Phase 2 — Complete
End-to-end prototype built
Calliope running as a full pipeline from substrate retrieval through articulation.
-
Phase 3 — In progress
Benchmarking against frontier models
Structured comparison across reasoning, grounding, and output quality. Early results on the research page.
-
Phase 4 — Planned
Closed early access
A limited group of evaluators will get hands-on access ahead of any broader release. Placeholder — date TBD
The broader initiative.
Ancilla means handmaid. Project Ancilla is the initiative Calliope sits inside — our closed effort, run with a small circle of partners and investors, to build assistants that are helpful by default, honest by construction, and bound not to cause harm.
Calliope answers "does this architecture work?" Project Ancilla is the open-ended question behind it: how far can it be pushed — and what does it take to make an assistant built this way dependable at scale?
— Ioma Labs, internal framingWant early access, or want to build it with us?
We're not public yet. If you'd like to be considered for closed testing, or you want to help build the next phase of Project Ancilla, reach out.