01
Context. Local.
The organization's own knowledge, terminology, evidence, and operating history — retrievable by the people permitted to see it, held on the appliance rather than in a vendor account.
01 / CobyAI II // Domatrix Series
Your terminal is the server.
Close the laptop. Keep the intelligence.
A portable, air-gapped AI server that gives your entire organization unlimited seats — running Context and Consensus on hardware you own and can carry out of the building.
Air-gapped or cloud-gapped · Unlimited internal seats · Three independent analytical paths
Configured build $62,995 · 60% deposit initiates
02 / The premise
Stop buying seats. Buy the server.
One appliance. Every authorized person in the organization. No per-user meter, no monthly renewal, no context living in someone else's account.
04 / Problem architecture
The problem is not access. It is context and control.
Everyone already has access to a model. What the team does not have is a place where the organization's own context lives, where the work can be inspected, and where the environment does not reset the moment a subscription lapses.
The portable Domatrix keeps the operating context next to the people responsible for the result — and keeps the analytical path visible from question to decision.
- STOP REPEATINGKeep the working context available to the people who are allowed to use it.
- SEPARATE THE PATHSRun independent analysis before the room settles on an answer.
- RETAIN THE RECORDPreserve sources, differences, decisions, and human review.
05 / The two capabilities
Context and Consensus. Everything else is hardware.
Two capabilities define the appliance. Context is what the organization knows. Consensus is how the system disagrees with itself before you are asked to decide.
02
Consensus. Reviewable.
Three independent analytical paths, compared rather than averaged. Agreement is reported with its dependence and its unresolved dissent intact — not converted into a claim of truth.
03
Authority. Human.
The system accelerates retrieval and analysis. It does not take the decision. Review stays where the responsibility already sits.
06 / System architecture
A laptop becomes a server. Governed.
The appliance serves the room. One machine, configured to the ceiling of the platform, running the models, the knowledge layer, and the review paths for every authorized user on the network it is permitted to reach.
06.01Local computeCore work stays in a configured local environment instead of leaving the building with every question.
06.02Shared roomServe every authorized internal user from one governed appliance and one knowledge environment.
06.03Reviewable pathsCompare independent work, challenge the evidence, and keep the route to the result.
07 / Node roles
Three paths. One review. No shortcuts.
The value of a third path is not that it votes. It is that it can reach a different conclusion, using a different route, and force that difference into the open where a person can weigh it.
A
Analyze
// NODE / INDEPENDENT WORK
B
Challenge
// NODE / COMPARATIVE WORK
C
Synthesize
// NODE / REVIEWABLE RESULT
08 / Model architecture
Put more models in the room. Give each one a job.
The model layer is configured around the work. Each family carries a defined role in the analytical path rather than sitting in an undifferentiated list of tools — because different lineages are what make the paths genuinely independent.
// 001
Google Gemma. Local.
Selected model family for controlled local workloads.
// 002
Mistral. Open.
Configured for the analytical role the work requires.
// 003
Llama 3. Ready.
A model family available within the configured environment.
// 004
DeepSeek. Focused.
Another analytical path available as configured.
// 005
Qwen. Selected.
Use the model family that fits the evidence and task.
// 006
NVIDIA acceleration. Fast.
Hardware acceleration for the workloads the appliance is built to carry.
09 / Knowledge architecture
Start with your intelligence. Then expand.
The base appliance can be expanded with specialty knowledge packages appropriate to the approved use case. Start with the institution's own intelligence, then add the specialty knowledge the institution actually needs.
Exact contents, licensing, update cadence, validation, provenance, and deployment method are established during implementation. No package is implied to be bundled, licensed, current, or automatically installed by this page.
Clinical and biomedical.
Literature, trials, terminology, guidelines, institutional protocols, and approved reference material.
Research and academic.
Scholarly metadata, institutional repositories, papers, datasets, and research libraries.
Legal and regulatory.
Jurisdiction-specific statutes, regulations, agency materials, contracts, and approved matter knowledge.
Safety and governance.
Cybersecurity frameworks, AI risk materials, safety guidance, controls, audits, and SOPs.
Engineering and technical.
Vendor documentation, APIs, manuals, runbooks, architecture records, and incident knowledge.
Institutional memory.
People, roles, decisions, workflows, services, evidence, provenance, and customer-owned knowledge.
10 / Control architecture
Security by architecture. Not decoration.
Air-gapped when the deployment requires the strongest separation. Cloud-gapped when the team needs to share the environment without giving up control. The deployment pattern is a decision, not a compliance outcome.
Air-gapped operationUse the strongest separation when the deployment requires it.
Cloud-gapped sharingShare the configured environment with the team without giving up control.
Encrypted storageUse local encryption and hardware-backed security as part of the appliance.
Permissioned contextDefine which knowledge, memory, and selectors each user can access.
Enterprise controlsConfigure authentication, device management, recovery, and administration.
Human reviewKeep people responsible for interpretation, decisions, and release.
11 / Mission briefing
Give the team a machine. Keep the authority.
The Triple Tri-Node is built for teams that need more than a chat window. It gives the room local compute, selected models, controlled knowledge, and a repeatable path through the work.
That is the difference between borrowing intelligence and owning the environment where intelligence becomes useful.
STATUS: CONFIGURED
STATE: LOCAL SERVER
PATHS: THREE INDEPENDENT
SEATS: UNLIMITED INTERNAL
REVIEW: HUMAN
12 / Build path
From configured hardware. To useful.
Specify
Define the work, users, knowledge, and controls.
Configure
Select hardware, models, storage, and security.
Install
Build the local environment and approved knowledge layer.
Validate
Test access, retrieval, workloads, and review paths.
Deliver
Put the server in the room and keep improving it.
13 / Straight answers
The questions that come up in the room.
Unlimited seats — what is the actual limit?
The appliance serves the authorized internal users on the network it is permitted to reach; there is no per-seat licence and no monthly per-user meter. The practical ceiling is concurrency against the configured hardware, which is established during specification against your actual workload.
Is air-gapped the same as compliant?
No, and we say so on every page. Air gap and cloud gap are deployment patterns, not regulatory outcomes. Local operation means the equipment is under your physical or administrative control — it does not by itself mean compliant, secure, or isolated. Controls belong in a validated deployment plan.
Why three paths instead of one good model?
Because a single confident answer hides the disagreement that mattered. Three differentiated paths can reach different conclusions and surface that difference for review. The caveat is real and we publish it: paths only add information when they differ in data, method, and evidence access — repeated prompts to related models do not establish independence.
What happens when the hardware ages out?
It gets replaced. The silicon is replaceable; the system is the product. The knowledge layer, permissions, provenance, and operating context are designed to carry across the change.
What does $62,995 include, and what starts the build?
The configured Triple Tri-Node Portable Domatrix at current product price. A 60% deposit initiates the build, configuration, and delivery path. Specification and validation happen before the appliance goes into operating use.
Your terminal is the server.
Own your future.
// 14 / PORTABLE DOMATRIX PRINCIPLE
15 / Commercial model
Own the portable server. Now.
$62,995 · 60% deposit initiates the buildBuild the configured CobyAI II Triple Tri-Node Portable Domatrix. The deposit starts the build, configuration, and delivery path.
// AIR-GAPPED OR CLOUD-GAPPED / UNLIMITED INTERNAL SEATS / HUMAN REVIEW
16 / Relevant reading
Read the evidence. Then build.
We publish the limits of our own architecture. Both papers below name Triple Tri-Node directly — and decline to make performance claims for it.
When Consensus Repeats Itself
Aggregation improves performance when component competence and error diversity are adequate — but count alone cannot substitute for independent information. Repeated consensus is neither intrinsically protective nor intrinsically harmful.
Knowing When to Stop
A defensible stopping framework weighs the marginal value of another differentiated analysis against cost, delay, dependence, and provenance loss. Bounded consensus is an adaptive decision process, not a fixed numerical recipe.
Measure First, Reason Second
Establish the measurement before the interpretation — and why the order matters when a system is asked to explain its own output.
Make the Operation Legible
The job is not to carry every answer. The job is to build a system that lets the answer be found, challenged, and reviewed.