Part V · Knowledge Evolution and Product Mechanics

Reference Journeys and Semantic-State Traces

End-to-end journeys test whether the whole system stays coherent across real work: thinking, workflows, editing, planning, delegation, recommendations, learning, external agents, team standards, and importing files.

Semantic situationSemantic custodyNegative authorityReconstruction

1. Why reference journeys matter

Component diagrams can conceal a broken system. A context compiler may work, a model may answer, a file may save, and an interface may render while the human purpose is still lost between them.

A reference journey treats the system as one temporal composition. Each journey declares:

the human purpose;
the initial semantic situation;
the cognitive movements required;
the source and context path;
the durable artifacts and records;
person-authority boundaries;
continuation and interruption behavior;
negative behavior that must not occur;
evidence required for acceptance.

These are research instruments as well as product specifications.

This chapter is the integration layer for the preceding architecture. It does not introduce another memory model, workflow system, or source of authority. It demonstrates whether the Fractal Seed, composed context, bounded judgment, governed memory, derived Atlas, exact effects, human authority, and reconstruction remain connected across time. A journey succeeds only when the whole sequence works.

2. A common state vocabulary

One scalar status cannot describe knowledge, work, authority, and effects at once. AIOS therefore uses independent state axes:

Axis	Representative states	Question answered
epistemic standing	observed → proposed → accepted or qualified → rejected, superseded, revoked, restored	What standing does this content have, for which scope?
work state	planned → active → produced or returned → integrated → completed; deferred or blocked	What stage has the work reached?
effect state	inert → previewed → authorized → executed → verified; failed or rolled back	What operational consequence has occurred?
resource availability	registered → eligible → selected → active; dormant or retired	Can this resource be resolved, and is it participating now?
custody and trust	local, imported, external, quarantined, reviewed	Which boundary has the object crossed, and how may it be used?

Namespaced notation makes the distinction explicit when ambiguity matters: epistemic.accepted, work.active, effect.authorized, or resource.dormant. The same object can occupy one state on every axis. A returned external report might be work.returned, effect.inert, epistemic.proposed, resource.registered, and custody.quarantined at the same time.

Several terms are not statuses:

canonical identifies an authoritative source location and version; it does not mean factually true or epistemically accepted;
registered means resolvable and inspectable; it does not mean active or mandatory;
integrated means landed in the correct artifact or ground; it does not grant epistemic standing;
authorized means an actor may perform a bounded effect; it does not accept the result as knowledge;
provenance establishes lineage; it does not establish truth;
receipt establishes what operation occurred; it does not establish quality or legitimacy;
reconstruction establishes recoverability of declared state; it does not establish the truth of stored claims.

The common journey envelope

Every journey instantiates the same whole-system movement:

human purpose
  → current semantic ground
    → bounded cognitive movement
      → composed context
        → semantic judgment
          → explicit standing
            → canonical target and version resolution
              → trusted preview and authorization
                → exact effect and post-state verification
                  → receipt and acceptance where required
                    → reintegration and renewed ground
                      → interruption and reconstruction test

The sequence can stop before effect, acceptance, or integration. What matters is that no skipped transition is silently assumed.

3. Journey A — thinking into durable composition

Purpose: A person begins with an ambiguous but consequential question and chooses to continue one conclusion into a durable document.

Thinking into durable composition

How an ambiguous question becomes a durable document only after the model returns a proposal and the person chooses what happens next.

sequenceDiagram participant P as Person participant T as Thinking movement participant C as Context composition participant M as Model judgment participant D as Durable artifact and companion participant U as Interface projection P->>T: State purpose, question, or tension T->>D: Save open working ground where required T->>C: Request bounded judgment from current ground C->>M: Coherent natural-language situation M-->>T: Judgment, rationale, uncertainty, continuations T->>D: Land inert semantic product D-->>U: Show ground and meaningful choices P->>T: Choose document continuation T->>D: Create or update canonical document and companion D-->>C: Publish renewed ground Note over T,U: Terminate and reload D-->>U: Reconstruct declared state, standing, lineage, and next movement

Semantic custody23-reference-journeys-and-semantic-state-traces--m01.mmd

The critical handoff comes after the model returns: its judgment is saved without changing any artifact and shown as a choice. Only the person’s continuation creates or updates the canonical document; after reload, its artifact and companion records must restore standing, lineage, and the next movement.

Required evidence: exact provider-visible ground; source path; semantic quality review; inert alternatives before selection; correct artifact and companion metadata; settled writes; interface meaning; reload equality.

Negative proof: no document is created merely because a model recommends it; no hidden session object is required after reload; no parser decides that the conclusion is true.

4. Journey B — a workflow that survives interruption

Purpose: A person accepts a useful structured process, completes several stages across sessions, and returns the result to the originating work.

A workflow that survives interruption

Workflow and recommendation states remain separate as staged work pauses, resumes from durable state, satisfies criteria, returns, and integrates.

stateDiagram-v2 [*] --> Recommended Recommended --> Dismissed: person dismisses Recommended --> Deferred: person defers Recommended --> RecommendationAccepted: person accepts relevance RecommendationAccepted --> LaunchAuthorized: person authorizes start LaunchAuthorized --> StageActive StageActive --> StageSaved: meaningful artifact saved StageSaved --> Paused: attention moves elsewhere Paused --> StageActive: reconstruct and resume StageSaved --> Evaluated: stage criteria reviewed Evaluated --> StageActive: revision required Evaluated --> NextStage: criteria satisfied NextStage --> StageActive NextStage --> Completed: final criteria satisfied Completed --> Returned: result routed to origin Returned --> Integrated Integrated --> [*]

Workflow23-reference-journeys-and-semantic-state-traces--m02.mmd

Acceptance of relevance and authorization to launch are separate transitions. Each stage saves a meaningful artifact before pause or advancement; evaluation may send it back for revision, and completion still requires return to and integration with the originating work.

The workflow does not own the person. The person can pause, revise, skip where the workflow permits, or leave the workflow. Recommendation standing and workflow execution remain separate: accepting that a workflow is appropriate does not itself launch it. Completion is tied to the outcome and artifact, not to clicking through stages.

Negative proof: a recommendation does not auto-launch; stage advancement does not occur before saves settle; interruption does not restart at stage one; completion does not create a duplicate artifact when same-file graduation is intended.

5. Journey C — structural editing without document damage

Purpose: A person asks the system to reorganize a complex document while preserving source meaning, voice, and unrelated material.

Structural editing without document damage

A whole-document reading leads to a bounded placement plan, authorized section edits, local and global review, and preserved lineage.

flowchart LR A["Read whole artifact and declared sources"] --> B["Identify purpose, argument, and constraints"] B --> C["Produce placement and transition plan"] C --> D["Person authorizes bounded edits"] D --> E["Apply section-level changes"] E --> F["Review local transitions"] F --> G["Review whole-document coherence"] G --> H{"Contract satisfied?"} H -->|No| I["Revise within scope"] I --> F H -->|Yes| J["Save revision, rationale, and lineage"]

Canonical artifact23-reference-journeys-and-semantic-state-traces--m03.mmd

Text equivalent

Read whole artifact and declared sources → Identify purpose, argument, and constraints; Identify purpose, argument, and constraints → Produce placement and transition plan; Produce placement and transition plan → Person authorizes bounded edits; Person authorizes bounded edits → Apply section-level changes; Apply section-level changes → Review local transitions; Review local transitions → Review whole-document coherence; Review whole-document coherence → Contract satisfied?; Revise within scope → Review local transitions.

The loop between whole-document review and scoped revision is the safeguard. Local polish is insufficient until transitions and the complete argument remain coherent, after which the revision, rationale, and source lineage are saved.

Negative proof: the system does not replace the whole document to solve a local placement problem; source passages do not lose attribution; polished transitions do not conceal conceptual contradiction; authorization scope is not expanded opportunistically.

6. Journey D — hierarchical planning and responsible replanning

Purpose: A long-horizon objective becomes an executable plan whose branches can change without losing the governing Why.

Hierarchical planning and responsible replanning

A governing purpose descends through outcomes, milestones, work packages, and tasks while new evidence can reopen any planning level.

flowchart TD W["Governing Why"] --> O["Target outcome"] O --> M1["Milestone or horizon A"] O --> M2["Milestone or horizon B"] M1 --> P1["Work package A1"] M1 --> P2["Work package A2"] M2 --> P3["Work package B1"] P1 --> T1["Bounded task"] P2 --> T2["Bounded task"] P3 --> T3["Bounded task"] E["New evidence"] --> R["Replanning review"] R --> W R --> M1 R --> M2

Plan23-reference-journeys-and-semantic-state-traces--m04.mmd

Text equivalent

Governing Why → Target outcome; Target outcome → Milestone or horizon A; Target outcome → Milestone or horizon B; Milestone or horizon A → Work package A1; Milestone or horizon A → Work package A2; Milestone or horizon B → Work package B1; Work package A1 → Bounded task; Work package A2 → Bounded task; Work package B1 → Bounded task; New evidence → Replanning review; Replanning review → Governing Why; Replanning review → Milestone or horizon A; Replanning review → Milestone or horizon B.

The vertical tree is not a one-way schedule. Every bounded task inherits why it matters, while the replanning path carries new evidence back to the governing Why and the affected levels of the plan, allowing one branch to change without being confused with the objective.

Every level expresses Why–How–What at its own resolution. A task knows what it produces and why; the plan can replace a branch without confusing the branch with the objective.

Negative proof: completed work does not force the future merely because effort was invested; a schedule does not become the source of purpose; local optimizations do not silently alter the governing outcome.

7. Journey E — delegated research and reintegration

Purpose: Several bounded roles investigate independent aspects of a question while the person remains free to continue other work.

Delegated research and reintegration

Differentiated commissions return evidence to a sufficiency review that compares, reconciles, or preserves disagreement before integration.

flowchart TD Q["Shared question and acceptance contract"] --> A1["Commission: empirical evidence"] Q --> A2["Commission: alternative model"] Q --> A3["Commission: failure and boundary review"] A1 --> R1["Evidence-bearing return"] A2 --> R2["Evidence-bearing return"] A3 --> R3["Evidence-bearing return"] R1 --> S["Sufficiency and independence review"] R2 --> S R3 --> S S --> C["Compare, reconcile, or preserve disagreement"] C --> I["Integrate into plan, artifact, or open question"] I --> N["Renewed context and next movement"]

Distributed cognition23-reference-journeys-and-semantic-state-traces--m05.mmd

Text equivalent

Shared question and acceptance contract → Commission: empirical evidence; Shared question and acceptance contract → Commission: alternative model; Shared question and acceptance contract → Commission: failure and boundary review; Commission: empirical evidence → Evidence-bearing return; Commission: alternative model → Evidence-bearing return; Commission: failure and boundary review → Evidence-bearing return; Evidence-bearing return → Sufficiency and independence review; Evidence-bearing return → Sufficiency and independence review; Evidence-bearing return → Sufficiency and independence review; Sufficiency and independence review → Compare, reconcile, or preserve disagreement; Compare, reconcile, or preserve disagreement → Integrate into plan, artifact, or open question; Integrate into plan, artifact, or open question → Renewed context and next movement.

The shared question branches by the kind of evidence sought—empirical evidence, an alternative model, and a failure review—not by agent count. The returns meet only for sufficiency and independence review, then may be reconciled or left in explicit disagreement before changing the receiving object.

Returns are not votes. Their sources, methods, scopes, and uncertainties determine how they can be combined.

Negative proof: the same source set repeated by three agents is not counted as three independent confirmations; a return is not accepted merely because it is fluent; background work does not capture the person’s foreground attention.

8. Journey F — recommendation, exploration, and chosen continuation

Purpose: The system notices several legitimate next movements without collapsing them into one optimization target.

Recommendation, exploration, and chosen continuation

How a small set of next movements can open direct action, review, or exploration while the person chooses what—if anything—to begin.

flowchart LR S["Current semantic situation"] --> R["Small contextual recommendation set"] R --> P1["Proceed directly"] R --> P2["Run a named review"] R --> P3["Explore adjacent possibilities"] P3 --> X["Explore map of workflows, roles, artifacts, and questions"] X --> Y["Preview why, effort, inputs, outputs, and effects"] Y --> Z{"Person selection"} P1 --> Z P2 --> Z Z --> A["Admit one bounded movement"] A --> G["Renewed ground; alternatives remain available or lapse"]

Semantic action field23-reference-journeys-and-semantic-state-traces--m06.mmd

Text equivalent

Current semantic situation → Small contextual recommendation set; Small contextual recommendation set → Proceed directly; Small contextual recommendation set → Run a named review; Small contextual recommendation set → Explore adjacent possibilities; Explore adjacent possibilities → Explore map of workflows, roles, artifacts, and questions; Explore map of workflows, roles, artifacts, and questions → Preview why, effort, inputs, outputs, and effects; Preview why, effort, inputs, outputs, and effects → Person selection; Proceed directly → Person selection; Run a named review → Person selection; Person selection → Admit one bounded movement; Admit one bounded movement → Renewed ground; alternatives remain available or lapse.

The exploration branch previews purpose, effort, inputs, outputs, and effects before selection. All paths converge on the person’s choice; only one bounded movement begins, while unchosen alternatives may remain available, lapse, or disappear when none is useful.

Negative proof: classification does not lock the person into one path; a menu item is not an effect; no recommendation is the correct output when no continuation is materially helpful.

9. Journey G — practice promotion and revalidation

Purpose: A repeated useful method becomes reusable intelligence without turning repetition into authority.

Practice promotion and revalidation

Observed practice becomes reusable only through evidence review, scoped acceptance, authorized canonical writing, registration, and later selection.

flowchart LR O["Observed use, outcome, or failure"] --> P["Proposed memory with source anchors"] P --> R["Evidence, counterevidence, scope, and trust review"] R --> H{"Accepted by the authority for this scope?"} H -->|"defer or reject"| X["Preserve disposition and lineage"] H -->|"yes"| A["Accepted derived knowledge"] A --> C["Exact authorized write to canonical files and metadata"] C --> G["Registered and eligible for fitting contexts"] G --> S{"Current occasion selects it?"} S -->|"yes"| U["Active in composed context"] S -->|"no"| G C --> Q["Qualified, superseded, revoked, or retired with lineage"]

Promotion23-reference-journeys-and-semantic-state-traces--m07.mmd

Text equivalent

Observed use, outcome, or failure → Proposed memory with source anchors; Proposed memory with source anchors → Evidence, counterevidence, scope, and trust review; Evidence, counterevidence, scope, and trust review → Accepted by the authority for this scope?; Accepted derived knowledge → Exact authorized write to canonical files and metadata; Exact authorized write to canonical files and metadata → Registered and eligible for fitting contexts; Registered and eligible for fitting contexts → Current occasion selects it?; Exact authorized write to canonical files and metadata → Qualified, superseded, revoked, or retired with lineage.

The sequence crosses several distinct boundaries: proposed memory, authority decision, exact write, eligibility, and activation. The side path preserves qualification, supersession, revocation, or retirement, while deferral and rejection retain disposition and lineage.

Promotion records occasion, contribution, counterevidence, misuse conditions, provenance, scope, acceptance, future-use evidence, and a review condition. Development progress, epistemic standing, registration, and activation are separate axes even when one interface presents them together.

Negative proof: frequency does not imply correctness; accepted guidance is not injected into every future context; a competing practice is not deleted to simplify the registry.

10. Journey H — correction after contradictory evidence

This journey is specified in detail in Self-Correction, Contradiction, and Epistemic Maintenance. Its invariant trace is:

accepted evidence-linked understanding
→ interruption
→ credible contradiction with source and trust standing
→ source and rationale recovery
→ scoped correction and affected-consumer analysis
→ appropriate authority accepts the correction
→ supersession without erasure
→ affected derived contexts and projections invalidated and recompiled
→ demonstrably changed later reasoning

Negative proof: metadata changes without changed reasoning do not count; historical rationale remains recoverable; semantic similarity alone does not authorize propagation.

11. Journey I — local system with an external agent environment

Purpose: A locally governed, self-contained domain knowledge system uses outside computational or agent capacity while minimizing disclosure and preserving local authority.

Local system with an external agent environment

Minimum-necessary local context crosses a bounded commission into external compute, while returned work stays quarantined until local acceptance.

flowchart LR L["Local purpose, artifacts, and authority"] --> M["Minimum-necessary context membrane"] M --> C["Bounded commission and permission contract"] C --> E["External agent or compute environment"] E --> R["Artifact, evidence, costs, and unresolved questions"] R --> Q["Local validation and quarantine"] Q --> H{"Person or delegated acceptance"} H -->|Reject| X["Retain or discard under policy"] H -->|Accept| I["Write accepted ground to local canonical artifacts"]

Local custody23-reference-journeys-and-semantic-state-traces--m08.mmd

Text equivalent

Local purpose, artifacts, and authority → Minimum-necessary context membrane; Minimum-necessary context membrane → Bounded commission and permission contract; Bounded commission and permission contract → External agent or compute environment; External agent or compute environment → Artifact, evidence, costs, and unresolved questions; Artifact, evidence, costs, and unresolved questions → Local validation and quarantine; Local validation and quarantine → Person or delegated acceptance.

The membrane controls what leaves the locally governed domain and what comes back. An external artifact carries evidence, cost, and unresolved questions into local validation; only person or delegated acceptance can write derived content into local canonical ground.

Negative proof: external completion cannot silently mutate accepted local ground; durable expertise does not become dependent on a provider’s hidden memory; sensitive source material is not sent merely because it is available locally.

The external return may be stored, indexed, and displayed while quarantined. Those operations preserve evidence and make review possible; they do not make the return governing memory. Only local acceptance for a declared scope can move derived content across that boundary.

12. Journey J — team standardization with local discretion

Purpose: An organization shares accepted workflows, definitions, and quality standards while individuals retain local projects, judgments, and controlled deviations.

Team standardization with local discretion

An organization package shapes team contexts and person-owned projects while local adaptations and objections return to organization review.

flowchart TD O["Organization-accepted package"] --> T1["Team context"] O --> T2["Team context"] T1 --> P1["Person-owned project ground"] T1 --> P2["Person-owned project ground"] T2 --> P3["Person-owned project ground"] P1 --> D1["Documented local adaptation"] P2 --> D2["Observed contribution or objection"] P3 --> D3["Observed contribution or objection"] D1 --> R["Organization review"] D2 --> R D3 --> R R --> U["Revise, narrow, preserve variance, or decline"] U --> O

Local knowledge-system fleet · canonical23-reference-journeys-and-semantic-state-traces--m09.mmd

Text equivalent

Organization-accepted package → Team context; Organization-accepted package → Team context; Team context → Person-owned project ground; Team context → Person-owned project ground; Team context → Person-owned project ground; Person-owned project ground → Documented local adaptation; Person-owned project ground → Observed contribution or objection; Person-owned project ground → Observed contribution or objection; Documented local adaptation → Organization review; Observed contribution or objection → Organization review; Observed contribution or objection → Organization review; Organization review → Revise, narrow, preserve variance, or decline; Revise, narrow, preserve variance, or decline → Organization-accepted package.

The downward path distributes shared operating ground without erasing local custody. The upward path treats adaptations and objections as evidence: organization review may revise, narrow, preserve variance, or decline rather than converting every exception into a new standard.

Standardization should define interoperable ground and acceptance criteria, not erase domain or practitioner judgment.

A signed or attestable package can establish origin, integrity, version, and declared owner. It does not establish local applicability, shared meaning, factual truth, regulatory certification, or permission to activate the package. Organization-level standing, team eligibility, local adoption, and permitted deviation remain distinct decisions.

13. Journey K — admitting existing work without automatic activation

Purpose: A person brings an existing draft or external source into a self-contained domain knowledge system while retaining provenance, local control, and separate judgments about relevance and standing.

Admitting existing work without automatic activation

An external file becomes addressable only through an authorized, verified custody operation and remains inactive until a later movement selects it.

flowchart LR E["External or unmanaged file"] --> P["Admission preview: object, source, destination, custody effect"] P --> A{"Person authorizes exact custody operation?"} A -->|No| N["No managed object created"] A -->|Yes| C["Copy or bind under declared policy"] C --> V["Verify identity, integrity, provenance, and companion metadata"] V --> R["Registered and addressable; inactive by default"] R --> S{"Later movement selects it as relevant?"} S -->|No| R S -->|Yes| X["Compose with visible provenance and standing"]

Artifact or source admission23-reference-journeys-and-semantic-state-traces--m10.mmd

Text equivalent

External or unmanaged file → Admission preview: object, source, destination, custody effect; Admission preview: object, source, destination, custody effect → Person authorizes exact custody operation?; Copy or bind under declared policy → Verify identity, integrity, provenance, and companion metadata; Verify identity, integrity, provenance, and companion metadata → Registered and addressable; inactive by default; Registered and addressable; inactive by default → Later movement selects it as relevant?.

The first decision governs whether any managed object is created. After the authorized custody action, identity, integrity, provenance, and companion metadata are verified; registration makes the object findable, but later use remains a separate judgment about relevance.

Does not establish Admission does not establish truth, acceptance, promotion, or activation; optional model orientation is separate and cannot authorize custody or future use.

For a batch, the preview makes the full intended consequence legible before authorization and defines whether failure is atomic or partial. Original preservation, copy-versus-bind policy, moved-file behavior, and local model or external-provider routing remain explicit rather than being inferred from file type.

Optional model orientation is a separate semantic movement. It may produce an inert summary, classification, or relationship proposal, but it does not authorize custody, establish the truth of claims, or activate the object in future context. If orientation crosses a device or provider boundary, the applicable disclosure and routing policy is part of the preview and receipt.

Required evidence: preview fidelity; exact authorization; source preservation; stable identity; provenance and custody state; companion metadata; explicit inactive state; restart reconstruction; correct later selection behavior.

Negative proof: admission does not imply acceptance, promotion, or activation; person authorship does not make every claim accepted; external origin does not make every file evidence; declining an admission does not conceal any already-performed disclosure or provider operation.

14. Cross-journey invariants

Every journey should preserve:

a visible human purpose;
explicit epistemic, intentional, and production ground;
one bounded cognitive movement at a time;
context composed for that movement;
explicit epistemic, work, effect, availability, and custody states where they differ;
structural validity separated from semantic judgment;
exact canonical target and version resolution;
a trusted preview of consequential effects;
person or valid delegated authority before consequence;
exact execution and post-state verification;
durable landing, provenance, and an operation receipt that is not mistaken for quality evidence;
acceptance separated from storage, indexing, registration, and integration;
honest failure and meaningful absence;
continuation into renewed ground;
reconstruction after interruption;
a negative-authority test;
proportionate evaluation of the actual outcome.

15. Journey evidence matrix

Journey	Semantic evidence	Durable evidence	Authority and effect evidence	Longitudinal evidence
thinking → document	judgment quality and context fidelity	artifact, companion metadata, and lineage	target resolution, explicit continuation, verified write	reload and resumption
workflow	stage-product quality	same artifact and stage state	separate recommendation acceptance, launch authorization, and verified advancement	pause, resume, return, and integration
structural editing	placement and coherence	revision and source lineage	resolved revision, computed delta, bounded approval, post-state check	whole-document re-review
planning	objective and branch alignment	plan revisions and work packages	authority for replanning and exact branch changes	project continuity after branch replacement
delegation	method and evidence sufficiency	commission, return, trust standing, and integration record	bounded commission and local integration decision	later use of integrated result
recommendation	relevance, rationale, and visible alternatives	origin-scoped proposal state	selection is distinct from effect authorization	alternatives expire, remain, or return appropriately
promotion	contribution, counterevidence, and scope	observation, proposed memory, acceptance decision, settled standing, and canonical lineage	named promotion authority and verified canonical write	selective reuse, qualification, and revalidation
artifact or source admission	object classification, provenance, and later relevance judgment	admitted object, stable identity, companion metadata, and custody state	full consequence preview, exact admission authorization, verified custody operation	inactive-by-default reconstruction and later context selection
correction	contradiction, trust standing, and corrected scope	immutable history plus current-governing lineage	correction authority, exact supersession, and recompilation evidence	changed later reasoning
external agents	minimum context and output quality	quarantine, source lineage, and local integration	disclosure eligibility, commission limits, and local acceptance	portability and provider exit
team standardization	shared purpose, evidence contract, and declared variation	package version, local adoption, deviations, and returns	organization standing separated from local eligibility and activation	revision, revocation, and preserved variance

16. Research use and connection

Reference journeys provide:

comparable experimental tasks;
trace formats for ablation studies;
interaction scripts for usability research;
boundary cases for security and governance review;
common objects for intelligence, interface, and implementation teams;
falsification criteria for broad system claims.

They turn “the system feels coherent” into observable sequences that can succeed or fail.

Current benchmarks increasingly evaluate time-extended state rather than isolated answer quality. MemoryAgentBench separates retrieval, learning, integration, and forgetting. τ-bench tests whether policy-guided actions reach the correct end state reliably across repeated trials. A 2026 matched-compute coordination study shows that decomposition can help or harm depending on task structure. Reference journeys extend this logic to the complete human–model–artifact system by testing context, standing, authority, durable landing, interruption, reintegration, negative behavior, and reconstruction as one observable sequence.

These studies do not validate the journeys or their state vocabulary. They clarify why an apparently successful response or tool trajectory is not enough. Deeper evidence is available in the relational-memory and AI-native architecture briefs.

17. What this page contributes to the whole

The reference journeys are where AIOS becomes falsifiable as an integrated product architecture. Each trace begins with human purpose, follows composed context into bounded judgment, separates standing from authority and execution, lands exact effects in local files, and tests whether the resulting state survives interruption and changes what happens next. Together they show how the publication’s upstream reasoning mechanisms produce downstream continuity, correction, collaboration, and control—without relocating intelligence or truth into one model, database, or visual map. Capability Horizons, Experiments, Proof, and Falsification turns these journeys into comparative tests of the integrated system.