🎥 The YouTube version is also available:
Explaining AI Decisions Through Evidence Combinations Chinoba’s Evidence Interaction
When AI provides answers or recommendations in business settings, it is important to show which materials support them.
Adding sources to RAG responses, displaying the referenced chunks, and making it possible to trace which regulations or procedure manuals were used are becoming basic requirements for trustworthy AI.
In practical decision-making, however, a conclusion is rarely determined by a single piece of evidence.
In maintenance decisions, equipment type, vibration values, temperature history, past replacement records, and exception conditions in procedures may only become meaningful when considered together. In quality judgments, not only inspection results but also the relationship to lots, process conditions, customer specifications, and change histories matters.
What is needed here is not merely an explanation that ranks “important evidence” from top to bottom.
Which combination of evidence supports this conclusion, and which missing conditions would cause the conclusion to change?
To answer this question, Chinoba can evolve the Evidence Graph into an Evidence Interaction Graph.
Attribution to a Single Piece of Evidence Is Not Enough
AI responses often involve multiple pieces of evidence.
For example, suppose an AI recommends that “this equipment should be escalated immediately for expert review.” Even if the screen displays the following sources, that alone does not provide a sufficient explanation:
- A procedure statement saying, “When temperatures are high, consider a shutdown decision”
- Recent temperature sensor values
- Historical failure records
- Time elapsed since component replacement
What truly matters is not which of these factors works independently.
A high temperature alone may simply reflect normal operating variation.
A long time since component replacement alone may not indicate immediate danger.
A similar failure in the past may involve different equipment.
However, when the equipment type matches, the temperature increase exceeds the specified range, the replacement cycle has elapsed, and the same signs appeared in past cases, the strength of the conclusion changes substantially.
These relationships can be called interactions among evidence.
Distinguishing Synergy from Redundancy
At least two important types of relationships exist among combinations of evidence.
Synergistic Evidence
These are pieces of evidence that may not be decisive individually, but strongly support a decision when combined.
| Evidence A | Evidence B | What the Combination Reveals |
|---|---|---|
| Equipment type | Vibration pattern | Consistency with a specific failure mode |
| Change history | Timing of quality deviation | A causal hypothesis between the change and the issue |
| Contract terms | Usage history | Whether special handling is permitted |
| Procedure exception | Measured value | A condition under which automated judgment should be stopped |
If synergistic evidence is overlooked, AI may fail to issue a necessary warning because “none of the individual documents contains a conclusive statement.”
Redundant Evidence
Redundant evidence consists of similar information sources that repeatedly support the same claim.
When multiple similar reports, duplicated FAQs, or several documents summarizing the same source material are retrieved, a larger number of references does not mean that the evidence is independently stronger.
If redundant evidence is presented in large numbers without distinction, AI may increase its apparent confidence, and users may mistakenly believe that multiple independent sources exist.
Chinoba must therefore not only identify synergy, but also detect duplication that comes from the same original source, the same observation, or the same prior judgment.
What Is an Evidence Interaction Graph?
A conventional Evidence Graph connects documents, chunks, rules, decisions, and actions as nodes.
An Evidence Interaction Graph additionally represents how a particular set of evidence influences a particular decision.
It can include relationships such as the following:
| Element | What Is Recorded |
|---|---|
| Decision or response | A conclusion or recommendation produced by AI or a human |
| Evidence Set | The set of chunks, data, rules, and observations involved in the decision |
| Interaction Type | Synergy, redundancy, contradiction, conditional reinforcement, alternative relationship |
| Applicability conditions | Target equipment, time period, process, authority, exception conditions |
| Degree of influence | An assessment of how much the decision changes |
| Verification status | Hypothesis, under validation, approved, withdrawn |
| Provenance | Original source, version, acquisition time, approving authority |
This allows AI explanations to move beyond simply saying, “I read this document.”
This recommendation was presented because the exception rule in Procedure Manual v4, the historical failure record for Equipment Type A, and the temperature increase over the past 24 hours are all present at the same time. A temperature increase alone would not trigger expert review.
With this explanation, users can understand not only why the conclusion was reached, but also which conditions would change the decision.
Counterfactual Receipt: Showing What Changes When Something Is Removed
A useful way to assess the importance of evidence is to ask counterfactual questions:
- Would the conclusion change if this evidence were absent?
- Would the same recommendation remain valid if this condition were different?
- Is one of these two pieces of evidence sufficient on its own?
- If evidence derived from the same original source were consolidated, would the conclusion remain unchanged?
In Chinoba, the results of these checks can be retained as a Counterfactual Receipt.
| Change Tested | Impact on the Conclusion | Interpretation |
|---|---|---|
| Remove the evidence of temperature increase | Escalation recommendation changes to pending review | An important necessary condition |
| Remove the historical failure record | Recommendation remains, but confidence declines | Reinforcing evidence |
| Remove the procedure exception condition | Recommendation changes to automatic shutdown | Important evidence defining a safety boundary |
| Remove multiple similar FAQs | Conclusion does not change | A redundant group of evidence |
A Counterfactual Receipt does not mean exhaustively testing every possible combination of evidence. As the number of evidence items increases, the number of possible combinations grows rapidly.
In practice, the following principles are useful:
- Prioritize validation of evidence groups estimated to have the greatest impact.
- Explore higher-order interactions by using relevant lower-order relationships as clues.
- Consolidate evidence derived from the same original source or observation, and remove redundancy first.
- Expand the scope of counterfactual checking for decisions involving safety, quality, or authority.
This is a design for finding the interactions necessary for explanation within limited validation cost.
Extending the Retrieval Receipt to Include Interactions
Chinoba’s Retrieval Receipt is a mechanism for recording which evidence was retrieved for which query, and why it was retrieved.
With the introduction of an Evidence Interaction Graph, the Receipt can be extended as follows:
| Receipt Layer | Question Recorded |
|---|---|
| Retrieval | Which evidence was retrieved? |
| Relevance | Why is it related to the current target or condition? |
| Interaction | With which other evidence does it become meaningful? |
| Counterfactual | Which removed evidence or condition would change the conclusion? |
| Decision | Who selected what, under which boundaries? |
With this structure, users can examine an AI-generated conclusion in ways that go beyond asking, “Are the cited materials correct?”
They can also ask:
- “Would the conclusion be the same without this material?”
- “Are the conditions supporting this conclusion actually satisfied in the current field environment?”
- “Are these multiple documents independent evidence, or repetitions of the same underlying information?”
This is the evolution from source display to decision-capable explanation.
Interactions Reveal the Structure of Organizational Knowledge, Not Only AI Reasoning
The value of an Evidence Interaction Graph is not limited to explaining AI responses.
It can also reveal the decision structures that an organization uses implicitly.
For example, if experienced practitioners always examine multiple conditions together when addressing a particular quality issue, that combination may represent important knowledge not explicitly written in the procedure manual.
Conversely, if frequently referenced materials are actually repetitions of the same original source, the organization’s knowledge foundation may be more fragile than it appears.
Interaction analysis makes it possible to ask questions such as:
- Which combinations of conditions create exception decisions?
- Which missing evidence would reduce the reproducibility of a decision?
- Which groups of materials are redundant and could be consolidated in update management?
- Where should additional observations or expert reviews be placed to improve decision safety most effectively?
These insights can be used to improve Semantic Chunk design, procedure manuals, priorities for data collection, and the division of roles between people and AI.
AI Accountability Cannot Be Measured by the Number of Citations
An answer with many citations may appear trustworthy at first glance. But what matters is not the number of citations.
What matters is how they are independent, reinforcing, contradictory, and supporting the conclusion under particular conditions.
A trustworthy AI explanation is not a list of evidence.
It is an explanation of the combination of evidence and the conditions under which the conclusion would no longer hold.
Chinoba preserves meaning and conditions through Semantic Chunks, connects provenance and relationships through an Evidence Graph, and records decisions and accountability through a Decision Trace. By adding an Evidence Interaction Graph and Counterfactual Receipt, it can evolve into a foundation that explains not only “why this conclusion could be made,” but also “what would need to be different for this conclusion to no longer be valid.”
Chinoba
Intelligence as Relationship
Research Platform
founded by
Masao Watanabe
AI Systems Architecture
Decision Trace
Human–AI Coordination
Algorithmic Governance
Related Research
This topic is part of the Chinoba Knowledge Base.
コメント