How Unison works¶
Intent becomes a governed outcome
Natural interaction, inspectable execution
Unison accepts a person's intent in the modality available to them, assembles a governed route through context, skills, tools, and models, and returns an experience suited to that person and situation.
Intent is the organizing unit¶
Intent describes the outcome a person is seeking together with the context that shapes it: identity, needs, preferences, timing, constraints, relationships, and the experience available in the moment. A spoken sentence, Braille command, gesture, image, device event, or structured workflow can all express intent.
The intent remains centered on the human objective as Unison selects context, policy, deterministic capabilities, governed inference, and a native output modality. This allows the route and expression to change while preserving what the person is trying to accomplish.
A request in motion¶
- Receive and interpret intentA local or remote channel captures speech, text, touch, Braille, visual input, device events, or another supported modality and maps it into a proposed human objective.
- Establish the person and contextThe trust layer resolves identity, assurance, active context spaces, consent, preferences, needs, and the policies that apply.
- Assemble a routeCore selects reliable workflows where they fit and composes governed inference, retrieval, or planning for novel parts of the request.
- Grant capabilitiesTools and connectors receive the data, credentials, purpose, and duration authorized for their portion of the work.
- Evaluate the outcomePolicy checks facts, sources, recipients, risk, confirmation, provenance, and recovery before an action or response proceeds.
- Compose the experienceThe semantic result becomes a native conversational, visual, Braille, tactile, or future-modality experience.
- Record useful evidenceA content-minimized receipt explains important decisions, actions, confirmations, and recovery paths.
Six cooperating responsibilities¶
Unison Core
Understands requests, assembles relevant context, plans work, and produces semantic outcomes.
Personal Data and Trust Store
Binds identity, governs private and shared context, brokers keys and credentials, evaluates policy, and retains audit evidence.
Capability Host
Contains tools and connectors within scoped grants for data, authority, purpose, and time.
Channel Gateway
Normalizes local and remote interaction, binds channel identities, detects replay, and applies assurance requirements.
Inference Broker and Runtime
Selects eligible local or remote models from signed manifests and task-specific constraints.
Unison Surface
Composes the same governed meaning through native experiences for each person, device, and modality.
The personal-appliance profile can combine modules into fewer containers while retaining their security controls and responsibilities.
Authority stays explicit¶
Identity authority belongs to the trust layer. Consent and policy authority belong to the governed personal node. Capabilities act through scoped grants. Models contribute interpretations, structures, descriptions, and proposals. The semantic outcome carries meaning, uncertainty, available actions, and recovery into each renderer.
This division supports replaceability. A new model, tool, channel, or modality adapter can enter through a versioned contract while the person's identity, memory, permissions, pending actions, and interaction preferences remain continuous.
Model routing as a governed capability¶
Each model version has a signed manifest describing its origin, interface, license, support state, and measured behavior. The Inference Broker evaluates the task, context policy, offline state, hardware, latency, quality, cost, and authorized disclosure before applying an inspectable ranking.
Candidate versions run synthetic semantic journeys in shadow and then progress through limited rollouts. Content-free health signals support rollback to a retained version. Exact facts and higher-risk actions receive deterministic validation and policy-controlled confirmation.
Capabilities can evolve through use¶
A novel request can begin with governed inference. Repetition can reveal stable structure such as familiar inputs, recurring tools, predictable policy checks, or a consistent output. Unison can turn that structure into a proposal with a clear purpose, permissions, tests, and evidence requirements.
The person reviews the proposal before activation. Approved capabilities are versioned and measured. Future requests can use the reliable route while the governed inference path remains available for changing circumstances.
Recovery is part of the result¶
Meaningful actions include confirmation, receipts, and recovery information appropriate to their risk. The system can explain the active person, context space, capability, recipient, important evidence, and available next action while retaining content-minimized operational logs.