Recommendation, Navigation, Menus, and Explore

1. Recommendation completes one movement and opens another

AIOS treats recommendation as part of the reasoning turn. Before the main response, the system orients to purpose and situation. The substantive movement then performs the needed judgment. Afterward, the system reads what formed and decides whether any continuation deserves to become visible.

The recommendation movement does not sit above the person's judgment. It reveals a small, meaningful horizon around the completed work. The person can accept an option, correct the orientation, continue in ordinary language, move elsewhere, or stop.

The diagram retains the implementation identifiers. The public account below uses the architectural phase names established in the Core Thesis.

2. Signals, focus, and action answer different questions

Context signals describe what is present. A turn can concern several domains, capabilities, subjects, output forms, audiences, channels, documents, and emerging terms at once. These observations help compose context and retrieve relevant ground.

Navigation focus names one optional branch organizing the immediate horizon. It gives the interface a present center without claiming to exhaust the meaning of a mixed situation.

Action selection identifies a consequence the person is choosing: continue a line, invoke an accepted workflow, open an Explore branch, develop an artifact, or review an integration.

The distinctions protect meaning. Observation enriches the situation. Focus organizes attention. Selection grants a bounded consequence. None substitutes for the others.

3. Menus create a shared local horizon

Menus let the model and the person work with the same meaningful possibilities while code retains exact identity. An established thinking movement can orient through controls such as:

Open is the uncategorized route in the top-level horizon. It preserves a way to continue or redirect without accepting a supplied category.

A framing judgment reads the situation and selects from the supplied positions. Deterministic mechanisms validate the selections and resolve the referenced guidance. The reasoning seat then receives that guidance in natural language, without the rejected options or menu machinery.

The model judges meaning; code resolves identity. This preserves semantic flexibility without asking software to infer cognitive posture from keywords.

4. Attention formation reads the Why

The attention-formation phase receives the person's current expression, the composed situation, available reasoning controls, the complete top-level navigation horizon including Open, controlled signal options, and any prior visible menu whose take-up needs interpretation.

It produces an opening read and selects the depth, aperture, ground, condition, and guidance appropriate to the turn. It can distinguish continuation from a shift or opening, identify an optional category, preserve multi-valued context signals, interpret whether the person took up a prior suggestion, and form one precise focus.

Attention formation establishes bearing. It does not decide which workflow will run, issue a command, choose a destination, or authorize an effect.

5. Bounded reasoning protects the substantive How

The bounded-reasoning phase receives the resolved orientation, admitted cross-domain ground, exact sources, and the person's current input. It performs the main cognitive movement and returns the person-facing reasoning, together with a concise public account of what the movement established when continuity requires it.

Recommendation machinery stays outside this context. Bounded reasoning does not need a list of every possible carrier or a generated Next Menu while it is trying to understand, discover, decide, compose, revise, or test. This separation keeps action bias from bending the substantive judgment toward whatever the system can offer next.

The concise account supports continuity. It records the public movement of the work rather than exposing or attempting to govern private chain of thought.

6. Integration and implication read the resulting What

The integration-and-implication phase receives the completed exchange, the resolved Why, relevant subcategories, controlled signal options, an exact snapshot of eligible actions, and top-level categories available for a possible shift.

It identifies what formed, whether a category or subcategory now fits, and whether the next context should continue, broaden, shift, or complete. It can carry forward supported signals, introduce newly visible ones, orient an integration review, curate up to four exact menu items, and write a public bridge.

Integration and implication keep three standings distinct. An eligible carrier is a factual possibility supplied by the system. A recommendation is a judgment that the carrier is useful here. Acceptance is the person's authorization to proceed.

7. The Next Menu presents several kinds of continuation

A concise menu can offer:

1. Continue thinking — develop one precise question, relation, or tension.
2. Develop document — form a named artifact around a clear contribution and ask.
3. Workflow — invoke one exact accepted workflow from the supplied eligible set.
4. Review integration — inspect a specific possible relationship or durable change.
5. Change context — move the next turn toward another exact branch without beginning work there.

Persistent routes remain available outside the generated menu: ordinary free text, continued thinking without selecting an action, and a change of direction. Recommendation broadens agency only when it remains one way to move rather than the interface's exclusive grammar.

8. Explore serves openness and bounded inquiry

Explore first names a person-facing space where a question, idea, or half-formed direction can develop without being forced into a deliverable or deadline. The system can ask discriminating questions, connect relevant ground, and make emerging directions or constraints visible.

Explore also names a delegated inquiry for an unknown too consequential or extensive for the active turn. Its directive establishes the exact uncertainty or alternative, why the return matters, accepted ground and source range, inquiry method, authority, boundaries, evidence-shaped return, and stopping condition.

The return lands at its originating object with provenance and limitations intact. The receiving judgment determines its bearing; the person or an existing grant determines consequence.

9. Internal capability becomes sparse public choice

An internal menu can be broad, technical, and explicit about recognition conditions, exceptions, and contraindications. It exists to help the system compose and evaluate possibilities.

A visible suggestion has a different obligation. It is plain, timely, specific, optional, and worth interrupting the person's attention to show. Many eligible internal possibilities may produce one public suggestion. They may also produce none.

This conversion from broad capability to selective visibility is a form of composition. It protects the person's attention while retaining a rich field behind the interface.

10. Exact compilation and semantic curation divide the work

The host builds an eligible horizon from known relationships: an explicitly selected continuation, active work, a prior capability, accepted inputs, authored workflow follow-ons, named integration possibilities, subcategory entry points, and category fallbacks.

This compilation answers what exists and can be resolved in the current snapshot. Integration and implication answer which of those possibilities matter now. The person answers what will happen.

Because exact identity is supplied, the model cannot make an unavailable workflow or destination real by naming it. Because usefulness remains a semantic judgment, deterministic code does not have to pretend that a keyword reveals the person's intention.

Software does not route complex work by keyword, and the model does not invent a carrier that the runtime cannot resolve.

11. Prior context changes bearing, not identity

Earlier work can make one branch more prominent and help the system preserve continuity. It does not erase Open, remove other categories, or establish a permanent classification of the person.

prior context
  → changes ordering, salience, and bearing
    ≠ fixes the meaning of the present turn
      ≠ deletes alternative directions

Recommendation state belongs to its originating file, project, or work context. When the person shifts domains, the local navigation history does not follow as a global identity profile.

12. Consequence advances through explicit standings

A model can produce a possible continuation. Integration and implication can observe it and decide it is worth recommending. The person can ignore, reject, park, or accept it. Acceptance grants a defined consequence; deterministic operations execute it; verification and integration establish what actually changed.

This lifecycle instantiates Chapter 10's standing law. Production and observation do not create a recommendation; recommendation does not create a grant; a grant does not create an effect; and an effect does not establish verified integration. Person acceptance here is take-up of a proposed consequence, while verified integration establishes the standing of what actually changed.

These transitions keep generative capability from becoming operational authority. They also preserve a truthful record: a compelling recommendation that was never accepted remains a recommendation, not unfinished work.

13. Silence is an intelligent recommendation state

Integration and implication return an empty menu when the response already completes the movement, the work has reached an honest stopping point, the eligible actions do not fit, uncertainty should stay open, another line deserves uninterrupted attention, or the person has already made the next movement clear.

The architecture treats this as success. A recommendation system is responsible for the quality of attention it protects, not the number of interactions it generates.

As Chapter 11 states, “Restraint is part of intelligence because attention has value beyond the system's opportunity to consume it.”

14. Correction renews local orientation

The person can reject the proposed category, select Open, ignore a suggestion, choose another workflow, continue in free text, clarify a signal, reopen an earlier branch, or state that no continuation is wanted.

Correction updates the relevant local situation and informs later composition. It preserves the prior record as history and does not generalize one correction into a permanent personal preference.

Boundary: recommendation reveals possibility; it does not exercise authority

Menus organize a local horizon. Signals describe context. Navigation establishes focus. Recommendations express judgment among exact eligible candidates. Only take-up or a prior bounded grant permits consequence. This separation is the basis of guidance without cages: the system can be genuinely helpful and directional while the person's freedom to interpret, redirect, and decline remains structurally real.

15. Failure conditions and evaluation

The principal failure conditions are distinct:

• Mixed-situation capture: one category is made to stand for a turn containing several live meanings.
• Automatic take-up: a recommendation starts work without person acceptance or a prior grant.
• Keyword routing: software treats surface terms as sufficient evidence of intention.
• Menu monopoly: generated choices become the interface's exclusive grammar.
• Engagement substitution: continued interaction replaces usefulness as the measure of recommendation quality.
• Global profiling: local navigation state becomes a durable classification of the person.
• Invented carrier: the model names a workflow or destination absent from the exact eligible snapshot.

Evaluation tests whether multi-valued signals preserve complex turns; whether optional focus improves orientation; whether separating attention formation, bounded reasoning, and integration and implication improves reasoning and reduces action bias; whether exact candidates resolve reliably; whether people understand recommendation versus execution; whether Open and free text remain natural; whether local state stays local; whether empty menus appear at appropriate moments; and whether Explore returns change their named receiving objects without expanding authority.

Recommendation gives the next movement a visible shape. Its achievement is not to keep the system moving. It is to help the person see what the present work has made possible and decide what deserves their attention now.