Versatil
The technology

A cognitive digital twin that runs a loop.

Under the surface: the metacognition brain. GraphRAG over a real graph, four memory tiers, bitemporal recall, conviction reasoning, and dissent kept first-class. Deterministic rendering, so the operator sees the real output, not an LLM in the middle.

Under the hood

Simple on top. Frontier underneath.

Everyone in the building gets an interface they can use on day one. The hard part, a graph, bitemporal memory, conviction reasoning, and a record that keeps dissent, runs underneath and stays out of the way.

SurfaceWhat anyone uses
Board, goals, approvals, chat. If you can run a project, you can run this. No prompt engineering, no new language to learn.
The brainWhat makes it different
Ontology graph, bitemporal memory, conviction scoring, dissent kept, and writeback when reality disagrees.
GroundWhere it runs
Your building, your keys, scoped ingest, no vendor-side copy. Sovereign by default, cloud when you choose it.
Our approach

Four things it does that a copilot does not.

A copilot answers. Versatil reasons, records the disagreement, and learns when reality proves it wrong.

01

Reasons, not just retrieves

It weighs the options and scores its conviction. Retrieval finds facts, conviction debates them and keeps the dissent.

02

Grounded in what you know

Every call starts from what the organization already knows, never a blank prompt.

03

Keeps the dissent

The position that lost stays legible, with the exact condition that would bring it back.

04

Learns when it is wrong

When a kept objection turns out right, the call reopens itself. The system sharpens instead of drifting.

See a decision land
app.versatil

Twin Activity: every decision with its conviction and its kept dissent, over a live ontology graph.

See the work-OS
The loop

It does not answer and forget. It runs a loop, and the loop writes back.

Stage 1 of 5
Ontology
The twin holds your organization as a live graph: people, decisions, sources, goals. Every call starts grounded in what is actually known, not in a blank prompt.
See the work-OS
The technology

A decision you can open, months later, and trust.

Under the marketing surface, every call is a record with its conviction, its sources, and the disagreement it kept. Nothing is averaged away. This is the layer a competitor cannot copy, because they threw the dissent out at write time.

Decision record, opened
specimen, not production data
Recorded

Vendor A for the ingestion pipeline.

DECISION-2026-Q1-014 ยท recorded 2026-02-11 09:42
conviction0.72
4 of 5 for, 1 dissent kept
opened from: decisions/2026-q1/ingestion-vendor.md
Dissent kept

Historian, 1 of 5. Vendor A's SLA has failed twice before in regulated orgs. The room chose speed. The disagreement was recorded, not averaged away.

reopen condition: incidents_per_quarter > 2
See it per profession
Full stack

Five layers, one brain.

Not a wrapper around a model. Each layer makes something implicit explicit, and writes it where the next decision can reach it.

Layer 1Ontology graph
Your organization as a live graph: people, decisions, sources, goals, and the edges between them. Every call starts grounded here.
Layer 2Bitemporal memory
What was believed, and when. Scrub to any instant and read the record as it stood, not as it was later rewritten.
Layer 3Conviction reasoning
Every call carries a score and a tally. The brain can tell you it already decided this, and decline to re-derive it.
Layer 4Dissent kept
The minority position is recorded with the exact condition that would reopen it. Never averaged away.
Layer 5Writeback
Reality comes back in. When a kept dissent's condition fires, the record reopens itself and the system sharpens.
Honesty

What is real, and what is being wired.

We do not say signed, sealed, or tamper-evident. We say inspectable, time-stamped, a record you can open.

Real today

The twin, kept

Ontology graph, bitemporal memory, conviction, and dissent kept are shipping mechanics, on specimen data you can inspect.

Being wired

Auto-reopen and push

Automatic reopen on a fired condition and push or stream legs are labeled being wired until they ship. No overclaim.

Never

Turnkey autonomy

The loop is gated, not a turnkey autonomous swarm. A human ratifies what matters. That is a feature, not a limit.

You do not need another pilot. You need enterprise AI you can answer for.

In production, on a sovereign twin, with the record of every consequential call. Early, with paying design partners, and honest about it.