Where expertise becomes a cognitive environment.
Every Cognitive Digital Twin we build starts the same way: we learn how your experts reason, then turn what we learn into an expert reasoning model your team can work inside. A pedagogical policy engine decides when to answer and when to challenge. Every answer traces back to a source you can check, whether a document or a dataset.
Pick the type that matches what you're trying to transfer.
Any type can be built for verbal or quantitative domains, from regulatory reasoning to signal diagnosis, and each comes with its own architecture, governance, and validation path. We choose the type with you before we build anything.
Domain Twin
Represents a domain rather than a person. Built from sources, cases, decision logs, and the structured reasoning of a community of practice. The easiest type to extend across teams.
e.g. Restructuring advisory · Threat modeling · Clinical workflow analysis
Composite expert Twin
Synthesizes the reasoning of multiple experts and primary sources. Keeps any one expert or school from dominating. Often the strongest fit for senior judgment.
e.g. Investment committee thinking · Cross-border compliance reasoning
Authorized expert Twin
Built on a real, consenting expert's method, clearly framed as “a system that simulates aspects of this person's reasoning, within these limits.” Governance is strict and licensing is explicit. The Twin never claims to be the person themselves.
e.g. A retiring partner's diligence method · A founder's product instinct
Institutional Twin
Learns how your organization decides: how you score risk, approve projects, run audits, develop people. Your house style of thinking, made teachable.
e.g. Your DD methodology · Your incident response playbook
Corpus Twin
Grounded in public or historical corpora. Useful for fields with deep written traditions or codified bodies of knowledge. Avoids individual impersonation entirely.
e.g. CCII reasoning · AI Act interpretation · Engineering standards
A first Twin, start to finish.
Choose one role where judgment is thinly held.
The best candidate is a role where a wrong call is expensive and only a few people make it well. Before we build anything, we agree on a baseline test with the AI switched off, so we can measure what changes.
Learn from the people who hold it.
We start with think-aloud sessions on real cases, then use cognitive task analysis and misconception mapping to surface what conversation alone won't. Every session runs with the expert's consent.
Build the Twin and its card.
We turn what we learned into an expert reasoning model and a pedagogical policy that decides when to help and when to hold back. A Twin Standard card documents it in writing, including what it cannot do.
Test the team with the Twin switched off.
The same people work the same kind of case, one they haven't seen before, without the Twin. That result, measured against the baseline, decides what happens next.
One Twin, four ways to work with it.
Learn mode holds three roles: the Mentor and the Sparring partner shape how your team practices, and the Examiner checks whether the learning stuck. Execute mode is for work where your team already knows the domain and needs a direct, cited answer.
Mentor
SCAFFOLDINGGuides a developing practitioner through unfamiliar territory step by step, asking orienting questions and offering hints rather than handing over the reasoning. The expert's full thinking is revealed at the moment the user needs it.
Sparring partner
ADVERSARIALChallenges the user's thinking with adversarial cases and counter-arguments an experienced colleague would raise. Productive friction by design. The system you turn to when you want to be wrong faster.
Examiner
VALIDATES UNDERSTANDINGTests transfer. Asks the user to teach back what they understood, and slips deliberate errors into the exchange for them to catch. Probes whether the principle was actually internalized. This is the role that closes the apprenticeship loop.
Execute mode
DIRECT ANSWERSDirect, cited answers for people who already know the domain and need a result now. Uncertainty is always declared, and there's a visible signal whenever you're trading learning for speed.
The moves that build cognitive autonomy.
The pedagogical policy engine orchestrates six specific moves. Each one is designed to keep your people doing the thinking at the moment when it's easiest to hand that thinking over.
First-you
Before the Twin gives its answer, it asks for yours, along with how confident you are in the call. It's designed to protect how you frame a problem and how you weigh evidence, two of the first skills to weaken once an AI starts answering for you.
Hint ladder
Help arrives in graded steps: a Socratic question, then a counter-example, then a principle, then a partial example. The full answer comes only when nothing less will do.
Adversarial cases
After a concept lands, the Twin introduces a realistic case that bends two of its conditions. Surfaces shallow understanding and builds the robustness no smooth explanation can.
Teach-back
The user explains the principle to a hypothetical junior, or to a deliberately fallible Twin. Catching the planted error is itself a learning signal. Teaching is the strongest test of having understood.
Transfer tasks
The same principle applied in a different domain, or with different stakeholders and different stakes. If a reasoning move only works in the case where it was learned, it wasn't really learned. Transfer is how we know.
Fading support
As your team's reasoning improves, the Twin does less. Hints arrive less often and cases get harder, until the system is actively working to make itself less needed. Anything else would be selling dependence.
Built for your domain.
What your experts teach stays on your systems. By default, so does every inference.
On-premise by default
Models run on your own systems, on standard GPU server hardware. We confirm the exact specification with you during scoping.
Pseudonymization gateway
Cloud inference is off by default. If you turn it on for a specific Twin, identifying data is pseudonymized before anything leaves your perimeter, and that Twin's card records it.
Air-gap ready
Internet-optional operation for air-gapped environments. You decide exactly what, if anything, may reach an external system.
Complete audit trail
Every inference is logged with its sources and reasoning chain, and available for review.
Start with a case that matters.
Bring a real problem from your domain. In a conversation with our engineers, we'll work through how a Twin would approach it.