Plain answers about Duplex.
The questions teams ask before they deploy.

With a pilot on your own work. Tell us where your team works and what Duplex should learn first. We agree the measures before it begins, and set Duplex up with you in the placement you choose.

On decisions and seats. A decision is one request Duplex takes part in, whether Duplex answers it or brings in your AI. Every quote comes in writing before you commit.

No. You bring your own model access and pay your provider for it, and Vareon adds no markup to tokens. A decision Duplex has learned needs no model call at all.

It stays in the placement you choose. In Vareon Cloud, Vareon runs Duplex for you and each customer stays separate. In your cloud, it stays under your controls. The one thing that leaves is a request to an outside model you connect, and that provider sees only that request.

Duplex is your AI model and an ACI model, working as one: runtime intelligence for your AI and your agents. Together they learn, plan, simulate, decide and act while the work happens, from one person to the whole enterprise. It learns from every result, plans, simulates, decides and acts within your rules, acquires skills of its own and carries its decisions and learning forward, wherever you work. What it builds is yours. Duplex is built on ACI, Adaptive Continual Intelligence: a new category of intelligence that learns, plans, simulates, decides and acts at runtime, in tandem with AI models.

Today’s AI is fragmented. The model, retrieval and application code each hold a piece of the intelligence, and none of them learns from the work. Duplex is the infrastructure layer that binds them into one system, and brings capabilities none of them has: it learns, plans, simulates, decides and acts while the work happens. It is a persistent learning runtime for AI agents: it keeps its context, enforces your rules and carries what it learned across sessions, agents and models. The centre of gravity moves from the model alone to the whole system, and what Duplex learns, plans, simulates and decides is yours. The model can change. Duplex carries the rest forward.

Anyone whose work runs on AI, from one person to the whole enterprise. One person starts in their own assistant, a team shares what it learns, a small business lets Vareon run it in Vareon Cloud, and a larger company brings it into every agent it builds through MCP + hooks, in its own cloud.

A harness gives your AI tools, prompts and steps, and does what it was set up to do. Duplex pairs your AI with a second intelligence, an ACI model, and the two work as one: it learns from every confirmed result, plans, simulates, decides and acts within your rules and carries your work across tools and models, and what it learns is yours. It works inside Claude, Codex, Cursor and Hermes, and inside your own agents through MCP + hooks.

94 capabilities in 17 groups: Context and observation · Rules and behavioural control · Tool execution · Learning and personalisation · Long-running tasks · Planning and decision support · Reusable skills · Persistent threads and sharing · Forgetting and data control · Status and explanations · Persistent intelligence · Personalisation · Multimodal intelligence · Confidence and autonomy · Runtime safety and oversight · Operational intelligence · Integration and portability.

No. Those store what happened and find what looks like it. Duplex is an intelligence that learns from experience, results and data at runtime: it understands the pattern behind each result, compresses it into what generalises, predicts what comes next and acts on it, within your rules. So it tells a look-alike apart, still knows a case asked in new words, follows a changed rule from the next request and gets faster with use. RAG, memory, vector databases, knowledge graphs and caches become tools it uses.

Your AI model is brilliant, but on its own at runtime: rented by the call, it starts every task from nothing, can’t learn from what your team confirms, loses context between sessions and agents, and nobody checks what your agents do before they do it. Duplex pairs it with a second intelligence, an ACI model, and together they learn, plan, simulate, decide and act at runtime: Duplex starts informed, learns from every result, chooses the route, plans the steps, simulates each one, checks each action against your rules, acquires skills and carries its decisions and learning across sessions, agents and models. Your AI gets what it lacks at runtime, the work gets faster and cheaper with use, and what Duplex learns is yours.

Your AI stays in place, and Duplex pairs it with an ACI model, a second intelligence beside it: the two work as one. Duplex answers, plans, simulates, decides and acts on what it has learned, in milliseconds, and brings in your AI for new and uncertain work with that learning attached. The confirmed result teaches Duplex for next time.

Two intelligences doing different jobs together, the way the brain does. The brain works with two systems: a slow one for general knowledge, and a fast one for each new experience. Your AI is the slow one. An ACI model is the fast one: it plans, simulates, decides and acts on what it learns. Together they are Duplex. Your AI provides general intelligence, learned slowly before deployment. The ACI model provides runtime intelligence: it learns fast from live work, plans, simulates, decides and acts on it. Cognitive science calls the pair complementary learning systems, and in Duplex the pair is the product.

Partly: Duplex is your AI model and an ACI model, working as one. The ACI model is a second model, an intelligence of its own, and it never replaces your AI. In Duplex, the ACI model works beside your AI: informed before the request, planning, simulating, deciding and acting during it, learning after the result, and carrying all of it to the next. A confirmed decision can run in milliseconds without a model call. New or uncertain work goes to your AI with Duplex’s runtime intelligence attached, and the confirmed result teaches Duplex for the next request.

Runtime intelligence: your AI model and an ACI model, working as one, learning, planning, simulating, deciding and acting while the work happens. Duplex learns from every result, chooses the route, plans, acquires skills, checks what your agents do, acts within your rules and carries it all forward. What it learns is yours. Two intelligences, one system: the combination is the product.

Anywhere your AI does real work: coding and desktop assistants, support and operations agents, enterprise workflows, high-volume AI products, teams adapting a model and regulated work. It is most useful where results can be confirmed, because Duplex applies each confirmed correction, decision or outcome to the next request it fits.

They become Duplex’s tools. Duplex calls each one when a step needs it, learns which one fits which situation and checks that what they return is current, sourced and dated.

ACI, Adaptive Continual Intelligence, is a new category of intelligence, created by Vareon. ACI is intelligence that keeps learning, planning, simulating, deciding and acting while your work happens. It carries everything it learned and decided forward, and works in tandem with the AI models you already use. ACI is the result of years of research into neural nets and their trilemma: plasticity, stability and editability. Duplex is built on it.

AI models learn before they are released, and then stay as they are. ACI learns, plans, simulates, decides and acts after: at runtime, from your work, for as long as it runs. It learns your work rather than the world, it builds on everything it learned before, and what it learns is yours. Intelligence that keeps learning, planning, simulating, deciding and acting at runtime, beside the model, is a category of its own.

In tandem. The model brings general intelligence: knowledge, language and reasoning. ACI brings what it learns from your work. It handles what it already knows, and brings in the model, briefed, for what is new. The model stays as it is, and you can switch models without losing what ACI learned.

AI that learns on the spot from one confirmed result, plans, simulates, decides and acts on what it learned within your rules, follows a new rule from the next request, gets faster and cheaper as it learns, knows when to bring in your model, handles cases it has never seen, acquires skills of its own, and belongs to you, wherever you work.

Until now, only the labs that train models could make AI learn. With ACI, every person and every company has AI that learns from their own work, and owns what it learns. What it learns compounds with every result, stays in the placement you choose, and is yours whichever model you use.

Duplex is built on ACI, and ACI is the intelligence in it: everything Duplex learns, plans, simulates, decides and does at runtime is ACI at work. Duplex is the product you use, inside the assistants your team works in and inside the agents your company builds, through MCP + hooks.

ACI is the result of years of research into neural nets and their trilemma: plasticity, stability and editability. Plasticity is learning the new, fast. Stability is keeping what it knows. Editability is changing what it learned. A neural net gets two at best: learn fast and it forgets, hold it steady and it stops learning, and what it learned is spread through all of it. ACI holds all three.

Faruk Guney built ACI, Duplex and his new startup on his own, after years of research into the trilemma of neural nets and more than fifteen years growing Norck into a global on-demand manufacturing and engineering company.

From results that show the work was right: a passed test, an approved review, an observed change or a later business outcome. It can also be pretrained on reviewed examples, past decisions, documents, rules and tool routes.

Yes. Pretrain Duplex on your reviewed examples, past decisions, documents, rules, tool routes and libraries of proven methods. It learns them before it goes live, so it is useful from the first request, and it keeps learning in production.

Your team sets the standard for each kind of work. Code may need tests and review, while a regulated action may need a person’s approval.

A correction belongs to the work area where it was made, and it helps later work that fits. Authorised users can review, correct or remove anything Duplex has learned.

Yes. Duplex learns the pattern, so it handles new wording and new details, and it spots a look-alike that needs a different answer. On misleading look-alikes it was wrong 7.5% of the time, against 45.3% for a cache.*

The new rule applies from the next request. Duplex replaces the outdated decision and leaves unrelated work in place. After a policy change it served 0% stale answers, against 72% for cache and retrieval.*

Continual learning, at runtime. Duplex learns from each result as it is confirmed, so the next request already benefits, with no training run and no GPU. Through a change, 100% of its earlier work stays intact.*

Ordinary use is the training: Duplex learns from every confirmed result and applies it from the next request. Pretraining gives it a head start, and your AI’s weights stay as they are.

For work that is new, broad or creative, and anything that needs fresh writing or reasoning. Duplex prepares exactly what your AI is given, with what it has learned attached. A decision Duplex has learned needs no model call.

If Duplex was pretrained on that kind of work, it answers from the start. Otherwise your AI takes it with Duplex’s knowledge attached, and once the result is confirmed, Duplex handles it next time.

Your AI, whenever the work needs new writing or open-ended reasoning, with what Duplex has learned attached. When Duplex has learned the work, it answers on its own.

Everything Duplex has learned lives beside your AI, so it moves with you to the new model. In the adaptation test, what Duplex learned moved to another model family with no retraining at all.

No, it does the opposite. What your organisation learns lives in Duplex, not inside one vendor’s weights, one chat transcript or one app’s memory. Change the model, the agent client or the provider, and the processes, rules, history and skills stay.

Yes. Your placement and settings decide which model is available and where each request goes.

Ordinary cloud servers, on CPU. Duplex needs no GPU of its own: a decision it has learned takes ≈97 ms, against ≈2,231 ms for a model call. Only your AI needs accelerators, and Duplex brings it in for new work.*

The goal, settled decisions, constraints and unfinished work continue within the same customer and work area. Unrelated work and credentials stay behind.

What you need to continue: goals, decisions and unfinished work. The full transcript stays where it was, and your team’s access rules still apply.

Yes, when the connections belong to the same customer and allowed work area. Identity, work-area and customer boundaries stay in place.

Yes, when both people are authorised for the shared work area. Personal or restricted work stays separate, under your access settings.

Context helps you continue unfinished work, and it can be temporary. What Duplex learns comes from confirmed results, and it helps every time it fits.

It applies confirmed corrections, updates decisions when a rule changes, carries active work forward, plans multi-step processes, uses your tools and cuts model calls on work it has learned.

Yes. Duplex learns how your processes respond, plans the shortest route and carries constraints and confirmed methods across steps and sessions. Your AI brings fresh reasoning to every unexpected turn.

Yes. It learns what actions change in your systems, predicts the effect of a step before it runs and proposes the shortest route to a goal. On a public enterprise benchmark, it beat three frontier models at predicting system changes. Every real step still runs through your controls.

Yes, once your team switches it on: give it a goal, the actions it may take and a way to check success. It explores safely, tests the skill before using it and removes it on request. It completed every task with the model switched off.

Yes. Duplex makes your model answer as if it had been trained on your work, and the model’s weights stay as they are. It beat fine-tuned adapters in 6 / 6 tests, with no training run, and the adaptation moved to another model family with no retraining.*

Yes. For well-defined tasks, Duplex writes a small program or reuses one that has already worked, tests it against known cases and runs it once it passes. Your team can pause or retire it at any time.

Yes. Documents, knowledge graphs and databases become Duplex’s knowledge sources, each with its owner, source and date. Duplex knows when a fact changes, and brings the right facts into each decision.

Yes. Duplex can share approved learning and context with every agent in the chosen work area. For anything it learned, you can see which agent it came from, with which model, and when. Each customer’s learning stays with that customer.

Your team decides what Duplex learns from, reviews what it learned and can correct or remove any of it. Your rules, approvals and controls apply to every action.

The four plugins are Claude, Codex, Cursor and Hermes. Teams that build their own agents connect them through MCP + hooks.

Four plugins are available today. If your team builds and runs the agent, MCP + hooks may fit; for any other assistant, tell us which one you use.

For the agents your company builds. They connect to Duplex through MCP, and the hooks bring Duplex in at every step: before each request, before each tool call and after each result. Each action is checked against your rules before it runs, with approvals where you want them and one history you can export.

A plugin works inside a supported assistant, through MCP: the assistant brings Duplex in at each step, and it builds your personal intelligence. MCP + hooks works inside the agents your company builds: the hooks bring Duplex in at every step of their work, and it builds your company intelligence. It is the same Duplex in both.

Before each request, before each model call, before each tool call, after each result, at each hand-off and at each approval. Duplex applies your rules first, then what it has learned, and each action is checked before it runs.

A plugin, when people work inside a supported assistant. MCP + hooks, when your company builds its own agents and Duplex should join every step.

Yes. Connected to the same work area, they share the same approved learning and context. Access stays limited to the people and agents you authorise.

A plugin or MCP + hooks, who shares what Duplex learns, and, if you like, the examples, decisions and documents to pretrain it on. Then the placement and the access rules that fit the work.

With a supported plugin, people keep working inside the assistant they already use. The team confirms good outcomes and corrects bad ones, and approval steps depend on the product and work area.

From the first request, when you pretrain it. A confirmed correction helps from the next matching case, and every confirmed result makes Duplex faster, cheaper and more accurate.

Yes, that is pretraining. Reviewed examples, past decisions, documents and rules make a strong start. Your team places each one in the right work area and can correct or remove it later.

Yes. In the console, authorised users can review, correct or remove anything Duplex learned, and see the history behind it. The exact controls depend on the product and deployment.

No. You set what matters: the goals, the rules and the actions that need your OK. Duplex handles the routine inside those rules and holds every agent to them. It checks each model call and tool action before it runs, asks you before the actions you choose, and carries what your team approved into every session and hand-off.

Yes. You can set up narrower work areas, and identity and access settings decide who shares each one. All of that learning stays within your organisation.

Duplex has two placements: Vareon Cloud, which Vareon runs for you, and your cloud, a cloud account you own. Your cloud is an enterprise solution. Every integration is available in both, subject to the deployment terms.

As an enterprise solution. Contact us at hello@vareon.com with the environment, where data must stay and your security requirements, and we plan the deployment with you.

Yes. There is a supported path between placements, and what Duplex learned moves with no retraining. Operating responsibilities and migration work are agreed for each deployment.

Yes. A deployment can use your own model access. Model-provider charges, credentials and the boundary they create are handled separately and set out in the quote and agreement.

Yes. Your inputs, reference material, history and everything Duplex learns from them serve your organisation alone. Work areas can add narrower boundaries for teams, products, regions or data classes.

Yes. Across six families of paraphrased attacks, it blocked every one, with 0.7% false refusals, the lowest of any approach that blocked them all.*

You choose who is authorised, in every placement. Vareon operates the hosted service, and has access to a deployment in your cloud only when you delegate it for an agreed purpose.

Your team signs in to the console through your own identity provider, traffic is encrypted in transit, and what Duplex learns and the history are encrypted at rest. Options to hold your own keys depend on the placement.

Yes. Authorised users can remove anything Duplex learned, restore an earlier version, and export what it learned with its history. The deletion and backup process for your deployment is covered in the security review and agreement.

Yes. Remove a result or an observation, and what Duplex learned from it is withdrawn with it, so nothing built on it lingers. The history shows the change.

Only what you allow. Personal data can be redacted before Duplex learns from it, sends it to a model or adds it to the history, and Duplex can apply what it learned without the original text.

It is append-only and tamper-evident: every change is added as a new entry, so an altered entry would show. Authorised users can export it.

Authorised users can remove its influence, and unrelated learning stays as it was. After deliberate poisoning, Duplex removed all 40 poisoned samples, and its legitimate learning stayed intact.

Retention is set per deployment. It depends on the data, placement, your configuration, backup design and the agreement that applies.

Duplex is in private pilots, and independent certification is planned. Until then, the security brief sets out the controls in place today for the placement you are reviewing.

The providers used for your deployment, and any subprocessor terms, are set out in your customer documents. In your cloud, you choose them.

11 results across cost, accuracy, continual learning, generalisation, planning, agents, security, new skills and adaptation, including three public benchmarks.

A cache replays an earlier answer when a request looks similar. Duplex learns the pattern behind confirmed results, follows rule changes from the next request and spots look-alikes, and it can still call your cache as a tool. After a rule change, Duplex held 92.5% accuracy where a cache fell to 42.6%.*

Keep it: retrieval becomes one of Duplex’s tools. Duplex calls it when a step needs a passage, learns which sources answer which questions and knows when a fact has changed.

It becomes a knowledge source for Duplex. Duplex reads linked facts when a decision needs them, knows when they change and learns which relationships matter for your work.

Yes. Fine-tuned and specialist models become tools Duplex routes work to, and Duplex can hand your training team confirmed examples when you want to train. Everything Duplex learns keeps working beside the new weights.

They become part of how Duplex acts. Duplex runs every model and tool call under your rules, and your guardrails and firewalls keep enforcing theirs.

An AI gateway routes calls and reports usage. Duplex is an intelligence: it learns from every confirmed result, plans, simulates, decides and acts on it, and through MCP + hooks it joins every step of the agents your company builds, with shared learning, continuity, approvals and one history.

Alongside them. Observability shows what your agents did and evals test it; Duplex learns from what your team confirms, so a correction applies from the next matching request. Your traces and eval suites keep working as they are.

Reflex is the runtime for any machine that decides and acts on its own, from edge devices to robots. It pairs your autonomy stack with an ACI model, on the device: every command passes the gate, and it learns, plans, simulates, decides and acts inside your rules, with or without a connection. Your stack is whatever runs the machine today: a VLA, an MPC, an autopilot, a PLC program or an AI agent. Your stack brings the skills it was built with. ACI brings the intelligence to learn on the spot, plan to the goal, simulate each move, decide in the moment and hold every command to your rules.

Five guarantees, on every machine, whatever runs it. Every command that acts passes the gate. Safety rules loosen only through an authorised, logged act. Every change is attributable and reversible. Learning stays on the device, bounded, with no network. “Unknown” means defer, not guess.

Whatever runs your machine today: a VLA, an MPC, an autopilot, a PLC program or an AI agent. It stays exactly as it is. Your autonomy stack and an ACI model, working as one: that is Reflex. The combination is the product. It starts from everything your stack can already do, and works with it, not in its place.

Both. Edge devices and robots: two shapes, one runtime. Edge devices are the controllers, gateways and smart cameras in plants, on sites and out in the field. Robots are the arms, mobile robots, drones, vehicles and vessels on land, in the air, at sea and under water, and the spacecraft in orbit.

It decides in the moment, on the device: what to do, which skill to run and when to defer. It decides whether each command may pass the gate, and whether it knows enough to act. It plans to the goal over what it has actually seen, chooses the skills that get there, and plans again when the world changes.

Yes. It simulates each move on the device before it makes it, and sees what a command will do before your machine carries it out. A move that would go wrong is changed on the device, before anything moves, and the move it makes still passes the gate.

Every command that acts passes the gate. Before anything moves, switches, opens or sends, Reflex checks it against your rules, and what fails does not run. It holds for every command, whatever runs the machine, and every decision at the gate is on record.

“Unknown” means defer, not guess. When it meets something it does not know, it hands the call to your stack or your operator. It never guesses. Every deferral teaches it: what your stack or your operator did next becomes something it knows, for the next time.

It learns fast and on the spot, on the device: one example or one deferral is enough, and the lesson carries to cases like it. It is focused learning, not general learning: it learns from what it is shown and from what it defers, inside the bounds you set, with no network and no training run. Your stack stays exactly as it is, and any lesson can be undone.

No. Learning never loosens a rule, and neither does time. Loosening one takes an authorised act, and that act is logged. Adding or tightening a rule takes effect from the next command, even while the machine runs.

Yes. Every change to a rule, a limit or what it has learned shows who or what made it, and when, and any of them can be undone: a lesson taken back exactly, or a machine returned to a point you saved. If a lesson ever cannot be taken back exactly, Reflex says so, rather than pretend.

No. It runs on the processors your machines already carry, on CPU, with no GPU and no cloud. It learns, plans, simulates, decides and acts with or without a connection: it works offline and air-gapped, never phones home, and its licence works offline too.

Yes. Each machine runs Reflex on the device and decides for itself, with or without a link. Across a fleet, each machine keeps its own rules, and skills move between them only when you choose. A skill learned on one machine can then go to others like it.

No. Reflex decides what the machine does next and holds every command to your rules at the gate. It sits above your certified safety controllers, interlocks and emergency stops, never in their place: they stay independent of it and keep the last word, and your team keeps each machine’s safety case, as it does today.

Both, alike, with the same five guarantees on every machine. It works with no connection, nothing has to leave the machine, and every decision is on record. In defence, that holds on kinetic and non-kinetic missions alike: every command passes the gate, within your rules of engagement.

Reflex is in beta. What it has shown so far comes from simulation and benchmarks, and we prove it with you, on your machines, in a pilot. Each pilot is planned with you, behind your own safety systems.

They are separate products, both built on ACI, Adaptive Continual Intelligence, and in each the combination is the product. Duplex is your AI and an ACI model, working as one, for your AI and your agents. Reflex is your autonomy stack and an ACI model, working as one, for any machine that decides and acts on its own.

The product, expected use, team size, placement and any dedicated capacity. Prices come in writing, with the assumptions behind them.

Duplex is a product of Vareon, Inc., the company behind vareon.com.

Use the contact page and mention the product, placement and kind of work you have in mind. The team will answer the question or point you to the page with the detail.