Exploring Long-Term Plans Without Losing Touch with Reality World Models, Virtual Decision States, and Evidence Sync in Chinoba

🎥 The YouTube version is also available:

AI for Long-Horizon Planning | World Models, Virtual Decision States, and Evidence Sync

Books: Al Coordination Engineering A Practical Guide: Designing Coordination, Autonomy, and Trust in the Multi-Agent Era

When AI supports business planning, it can often respond reasonably well to short questions.

“What are the key points of this procedure manual?”
“What are the possible causes of this anomaly?”
“Which materials should we review?”

However, once the task becomes an improvement plan spanning weeks or months, the introduction of a new business system, the prevention of recurring quality problems, or a transformation that crosses multiple departments, the difficulty rises sharply.

In long-term planning, it is not always sufficient to move in a straight line toward the goal. It may be necessary to begin with field observation. Safety checks may need to be completed first. Returning to a small prototype may prevent far greater rework later in the process.

For both AI and organizations, long-term planning is not simply a matter of choosing “the next best action.”

It is the problem of exploring multiple paths that include uncertain intermediate states, while repeatedly returning to executable decisions without drifting away from real-world evidence.

At Chinoba, this challenge can be approached as a design for long-horizon planning that combines a World Model, Virtual Decision States, and Evidence Sync.

Why Long-Term Planning Is Difficult

In short-term optimization, actions that reduce the problems currently visible are often effective.

Long-term planning, however, has a nonlinear structure.

Consider a manufacturing site seeking to reduce recurring quality defects. If the goal is only to lower the immediate defect rate, the organization may naturally lean toward measures such as increasing inspections or tightening judgment criteria.

In practice, however, a more effective path may be the following:

1. Examine the relationship between past defect records and process conditions.
2. Correct missing data and inconsistencies in recording practices.
3. Test a limited detection rule in one process.
4. Evaluate the confirmation burden on the field and the false-positive rate.
5. Incorporate the approach into approved operating rules and expand the scope of application.

At intermediate stages, actions that initially appear to be detours may be necessary.

For this reason, if AI is asked only to choose the action that seems closest to the final goal, it may skip essential preparation, validation, and consensus-building.

The farther ahead a plan attempts to look, the more assumptions and branches it creates, and the more easily it can be pulled by intermediate hypotheses. The more plausibly AI describes a detailed future, the more necessary it becomes to verify whether its assumptions still retain real-world evidence.

Viewing the World Model as an Organizational Change Model

The World Model referred to here is not a massive simulator intended to reproduce reality in its entirety.

In Chinoba, a World Model is an evidence-grounded model of change: a model that represents how business states may change when particular actions are taken.

For a quality-improvement initiative, for example, it can maintain relationships such as the following.

This model does not need to be built only from mathematical formulas. It can be constructed from Semantic Chunks, an Evidence Graph, business workflows, and past Decision Traces, as knowledge about how particular actions may affect particular outcomes under particular conditions.

The essential point is not to treat predictions as facts. Every prediction should carry its evidence, confidence level, applicability conditions, and unresolved issues.

Using Virtual Decision States: Exploring Without Directly Changing Reality

In long-term planning, it is necessary to compare multiple options. Yet the real operational state, formal rules, and approved knowledge must not be changed during exploration.

For this reason, Chinoba uses Virtual Decision States.

A Virtual Decision State is an exploration-only state that represents what may happen next if a particular action is selected.

Consider three possible improvement approaches:

• Option A: Immediately tighten inspection criteria.
• Option B: Prototype detection support in a limited process.
• Option C: First improve record fields and data quality.

For each option, AI can generate a Virtual Decision State and expand the necessary evidence, expected effects, risks, and conditions for the next validation step.

At this stage, none of these options is a formal decision. AI may compare them freely, but the real execution state does not change.

Virtual Decision States are scaffolding for exploring the middle of a long-term plan. Because they are not reality itself, they allow nonlinear paths, alternatives, and possible failures to be considered safely.

The Risk of Virtual States: Optimizing for a Plausible Future

A design based on Virtual Decision States has major advantages, but it also carries risks.

AI can construct plans by connecting intermediate states that have weak evidence but appear plausible. Even if assumptions gradually drift away from reality, each individual step may still appear internally consistent.

This can be understood as a form of overfitting to virtual states in long-term planning.

For example, AI may assume that improving data quality will produce highly accurate predictions, and then proceed to create staffing plans and operating designs based on those predictions. In reality, the necessary labels may not exist, the granularity of records may differ across processes, or the field may be unable to sustain ongoing data entry. These issues may only become visible later.

For this reason, Chinoba does not allow Virtual Decision States to be optimized freely. Instead, it establishes the following principles:

• State transitions without evidence must be explicitly marked as hypotheses.
• Proposed actions must not exceed approved constraints or human decision boundaries.
• Important intermediate states must be connected to an observation or validation plan.
• The higher the risk of an assumption, the earlier it should be checked against reality.

AI creativity should be used not as unsupported certainty, but as diversity in testable hypotheses.

Evidence Sync: Synchronizing Explored Plans with Reality

The key to handling long-term planning safely is to synchronize virtual exploration with real-world evidence on a regular basis.

At Chinoba, this can be called Evidence Sync.

Evidence Sync compares Virtual Decision States created during exploration against real evidence, the latest operational state, approved rules, and human review.

The comparison includes questions such as:

• Are the referenced documents and data still current and valid?
• Are the assumed conditions for application actually satisfied?
• Does new observation support or contradict the plan’s assumptions?
• Has the plan crossed a boundary that requires human approval?
• Are unresolved matters being treated as if they were facts?

As a result of synchronization, the plan can receive one of three treatments.

Through this mechanism, AI does not verify reality only once at the end of planning. It can return to reality repeatedly throughout the process.

Separating Freedom of Exploration from Strictness of Execution

In Chinoba, exploration states and execution states must be clearly separated.

During exploration, AI may generate a broad range of alternatives, examine failure paths, and compare hypotheses. In contrast, changes to external systems, external communications, rule revisions, equipment operations, and similar actions should be permitted only when they have a Decision Trace that satisfies the required evidence and approvals.

This separation allows AI to explore long-term possibilities boldly while behaving cautiously toward the real organization.

Long-Horizon Decision Trace: Making the Planning Process Itself an Asset

In long-term planning, the true value lies not only in the plan that is ultimately adopted.

It also lies in which hypotheses were tested, which evidence supported which assumptions, when the plan was revised, what was withdrawn, and why another path was selected.

If these are preserved as a Long-Horizon Decision Trace, the planning process itself becomes a knowledge asset for future improvement.

In the next improvement project, for example, the organization can reuse lessons such as: “Postponing the validation of record quality led to substantial rework after prototyping,” or “Early expert review helped define a safe validation scope sooner.” Such learning can be fed back into the next World Model and Decision Scheduler.

The important point is not to preserve only successful plans. Withdrawn options, paths that were not selected, and hypotheses disproved midway also hold value for making future exploration more intelligent.

Not AI That Thinks Longer, but AI That Can Keep Learning Longer

AI that is strong in long-term planning is not AI that can accurately predict the distant future in a single attempt.

It is AI that can generate hypotheses, explore intermediate states, synchronize with evidence, revise plans when necessary, and carry its decision history forward to the next challenge.

Long-term intelligence is not the ability to fix the future.
It is the ability to explore an uncertain future while continuously returning to real-world evidence.

Chinoba constructs a World Model as an organizational change model, safely explores plans through Virtual Decision States, and compares them with reality through Evidence Sync. Through Long-Horizon Decision Trace, it applies the trial and error of planning to future decisions.

In this way, AI can move beyond being an entity that simply generates planning documents. It can become a knowledge and decision-making foundation that helps organizations pursue long-term change without losing their grounding in evidence.

Chinoba

Chinoba

Intelligence as Relationship

Research Platform

founded by

Masao Watanabe

AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance