A Cognitive Digital Twin is an extension of how you think.
The Japanese call it jinba ittai, horse and rider as one body: the horse reads the archer's intent before it turns into motion, so the archer can simply shoot. That is the quality we engineer. A Cognitive Digital Twin understands what you are reasoning about and steps back the moment you've found your footing, so the joy of working precisely as you intend is within reach of everyone on your team, including people who haven't spent twenty years earning it. The pedagogy underneath the language model is the hard technical problem.
What sits underneath every Cognitive Digital Twin.
The language model is the least defensible part. Anyone can swap it. The lasting value is in four assets we build with your experts and have validated by experts outside the build team, and they keep improving with use. They are layers 03, 04, 05, and 07 in the architecture below.
Expert reasoning model
Expert reasoning is the work an expert does before they answer. It shows up in the questions they ask first and the signals they treat as diagnostic in a document or a dataset. It also shows up in what they choose to ignore, and in the shortcuts and trade-offs they've learned to treat as legitimate. We bring all of that to light by working alongside them. Documents, dashboards, and models tell you what experts know. The expert reasoning model shows how they think while they work.
Misconception model
Every domain has its characteristic mistakes, including some that mislead even careful people. We map these systematically alongside the expert reasoning, so the Twin can diagnose a developing thinker's error and respond with the precise intervention an experienced mentor would offer.
Pedagogical policy engine
The deepest moat. The policy engine decides whether to answer or hold back, and whether the user needs help or a little more struggle. It also decides when to surface an adversarial case rather than take the user's response at face value. Knowing when to stay quiet is what makes the Twin an environment for building competence.
Validation loop
The metric that matters is how well your team reasons when the AI is turned off. We measure it with structured pre and post evaluations, retention checks, and transfer cases the team has not seen. If that unaided performance is not improving, the system is not working, no matter how much it gets used.
How the seven layers fit together.
A Cognitive Digital Twin is a layered system where each component does a specific job, and the policy engine decides how they cooperate. The layer that makes it different from a chatbot with documents attached is layer five, the one that governs when not to answer.
Domain ontology
A navigable map of the concepts, sources, controversies, and paradigmatic cases that organize your domain. It represents how the domain itself sees the world.
Task grammar
The recurring cognitive operations of the domain: diagnose, classify, design, evaluate, prioritize, decide under uncertainty. The Twin knows which task is in front of the user, and what mental moves it requires.
Expert reasoning model
The structured representation of how your senior experts actually work a problem: the first questions they ask, the signals they treat as diagnostic, the thresholds at which they decide, the criteria that would change their mind. It's the part of the system that's uniquely yours.
Misconception model
The characteristic errors of newcomers, from confused concepts to shortcuts that look plausible until they fail. Each error is mapped to a corrective intervention and to a contrast case that surfaces it.
Pedagogical policy engine
The orchestration layer. It decides whether the next move is a question or a hint, and whether this is the moment for an adversarial case rather than a quiet pause. Sometimes the right call is to step back entirely. Every interaction is shaped by these decisions, more than by anything the language model defaults to.
Minimal learner state
A lightweight, task-bounded profile: which concepts the user has mastered or still finds fragile, and where their confidence is miscalibrated. It adapts the Twin to each person.
Validation & uncertainty loop
Continuous measurement of domain accuracy, fidelity to expert reasoning, pedagogical effectiveness, and unaided performance improvement in the people using the system. The last metric is the one that matters most.
Learn when you want to grow. Execute when you need to ship.
Speed and learning are a real trade-off. We make it visible: the user picks the mode, and the system marks clearly when a choice has a cognitive cost.
Learn mode
BUILDS COMPETENCEThe default for new domains and developing practitioners. The Twin asks before it answers and surfaces your assumptions. Then it runs adversarial cases and asks you to teach back what you understood. Support fades as your competence grows. It is slower in the moment, and the gains build over time.
- ›First-you: state your hypothesis before you see the expert's
- ›Hint ladder: progressively richer help, only when needed
- ›Adversarial cases: realistic edge cases that stress your reasoning
- ›Teach-back: explain what you learned, with intentional errors to catch
- ›Transfer tasks: same principle in a different domain. Does it still work?
- ›Fading support: as you improve, the Twin does less
Execute mode
PRODUCES OUTPUTFor the operational tasks where you already understand the domain and need a result. The Twin gives direct answers, and drafts documents or summarizes evidence when that's what the task needs. Every output includes full citations and declared uncertainty. A clear signal also shows whenever you're trading learning for speed.
- ›Direct answers with grounded citations
- ›Document drafting, scenario modeling, structured summaries
- ›Confidence and uncertainty made explicit on every output
- ›One-click pivot back to Learn mode whenever you want to go deeper
Every Twin ships with a Twin Standard card. Limits included.
Every Twin Standard card declares in writing the Twin's sources, the experts involved, its validation status, and its limits, including what the Twin will not attempt. The example below is set in financial restructuring.
Restructuring advisory reasoning Twin
Illustrative exampleCrisis management, debt restructuring, viability assessment
Senior associates · directors · investment committees
- ›CCII as currently in force, and composizione negoziata practice
- ›Bank of Italy supervisory notes (2018–2026)
- ›Anonymized historical mandates (n=87)
- ›Decision memos and post-mortems from your firm
- ›3 partners (avg. 24 years of experience)
- ›2 external court-appointed administrators
- ›1 lender-side credit officer
- ›Diagnose viability vs. liquidation across scenarios
- ›Frame restructuring options against creditor classes
- ›Pressure-test management plans for hidden assumptions
- ›Coach junior analysts through unfamiliar sectors
- ›Does not provide legal advice; escalates to counsel
- ›Does not value specific securities
- ›Not validated on cross-border insolvency outside the EU
Expert review · before/after test with the Twin switched off · transfer tasks on novel cases
2026-04-10
How we build systems that earn trust.
Every answer traces to a source.
Each answer shows the sources it rests on, and the reasoning behind it is logged for review. In regulated industries, an articulate hallucination is worse than silence.
Replacement is the wrong target.
We work alongside your experts to make their reasoning available across the organization, and to develop the next generation of them.
Friction as a feature.
Productive difficulty is how people build competence, so we engineer the right amount of friction into the right moments. The Twin makes you sharper by asking a little of you.
Built to run, not to demo.
We optimize for production reliability in regulated environments: continuous monitoring and graceful failure, backed by expert review on a schedule. That is what it takes for a system to be trusted with real work.
Security built into the architecture.
What your experts teach stays on your systems. By default, so does every inference.
Built for GDPR
By default, processing stays inside your perimeter. That keeps personal data under your control and makes your GDPR obligations easier to meet.
On-premise by default
Models run on your hardware, inside your infrastructure. Cloud inference stays off unless you turn it on for a specific Twin.
Complete audit trail
Every inference is logged with its sources and its reasoning chain.
Air-gap ready
Deployable on networks with no internet connection, for environments that require it.
A working session, with your problem.
Our engineers will walk you through the architecture behind a Cognitive Digital Twin, using a real challenge from your domain. By the end, you'll know whether it's the right answer for you.