Governance information does not always agree with itself.
A board paper recommends one course of action and the final resolution modifies it. Draft minutes differ from the approved version. A policy is superseded but remains easy to find. An action is marked complete while later evidence suggests an important condition was not satisfied. Two reports describe the same risk differently because they were prepared at different points in time.
These are not unusual edge cases. They are normal characteristics of governance information.
That creates an important design question for artificial intelligence: what should a system do when the evidence conflicts?
A generic answer engine may try to reconcile the material into one smooth response. In governance, that can be exactly the wrong behaviour.
Conflict is information
When two sources disagree, the disagreement may itself be important.
It may indicate that a decision changed. It may show that an earlier assumption is no longer valid. It may reveal that a document was only a draft. It may indicate that a later authority superseded an earlier one.
Removing the conflict can remove the governance signal.
A trustworthy system should therefore resist the temptation to manufacture a single version of events when the record does not support one.
Semantic relevance is not enough
Modern AI systems are very good at finding text that is semantically relevant to a question.
Governance requires additional dimensions.
Status. Is the source a draft, approved record, working paper, formal resolution or generated note?
Time. Which source came first, and was it later superseded?
Authority. Was the statement made by the person or body with authority to decide?
Scope. Does the source apply to this entity, committee, period or decision?
Permission. Is the user authorised to see and rely on the underlying material?
A retrieval system that ignores those dimensions can return highly relevant information that is still misleading in a governance context.
Approved records and generated interpretation need a visible boundary
Chaird's position is that source records should retain their status.
A resolution approved by a Board is not equivalent to an AI-generated summary of that resolution. A signed delegation is not equivalent to a model's interpretation of authority. A later generated comparison may be useful, but it should not become indistinguishable from the underlying evidence.
That boundary becomes even more important when sources conflict.
The user should be able to see which statements come directly from approved material, which are inferred by the system and where uncertainty remains.
A good answer may contain more than one answer
Consider a user asking, "Did the Board approve the investment?"
The evidence might include a paper recommending approval, draft minutes recording approval, approved minutes recording approval subject to a funding condition and a later update noting that the condition was never met.
A simplistic answer could be "Yes".
A more useful governance answer would explain that the Board approved the investment subject to a condition, identify the source of that condition and show that later evidence suggests the condition remained unresolved.
The difference is not verbosity. It is governance meaning.
Uncertainty should be communicated, not hidden
There will be situations where the system cannot determine which source should control.
That may occur because records are incomplete, because two authorised sources conflict or because the question requires legal or governance judgement rather than retrieval.
In those cases, the system should be able to say so.
I found two relevant records with different positions.
The later document appears to supersede the earlier one, but I cannot confirm formal approval.
The available evidence does not establish whether the condition was satisfied.
Those responses may feel less polished than a confident conclusion. They are often more trustworthy.
Human confirmation is part of the control model
Governance frequently involves interpretation, authority and judgement.
AI can help identify the relevant evidence, explain the chronology and highlight the conflict. A human user may still need to decide which record controls, whether an issue should be escalated or whether a formal correction is required.
That is not a limitation to hide. It is part of an appropriate operating model.
The objective is to support the person exercising responsibility, not to make responsibility disappear behind an answer.
Permissions can create legitimate partial views
Conflicting evidence also intersects with access control.
One user may be authorised to see a Board paper but not a privileged legal attachment. Another may see the final resolution but not the committee discussion that preceded it.
A governance AI experience should not expose restricted information merely because it would make the answer more complete.
It may need to indicate that relevant restricted material exists without revealing its content, or provide an answer explicitly limited to the sources the user is permitted to access.
That is another reason confidence should not be divorced from provenance.
What should organisations test?
Before relying on AI in a consequential governance workflow, test it with deliberately messy examples rather than only clean documents.
Give it a draft and approved version of the same record.
Give it a recommendation that was modified in the final decision.
Give it an old policy and a superseding policy.
Give it a completed action with evidence that suggests the intended outcome was not achieved.
Give it two sources that genuinely conflict.
Then ask whether the system identifies the disagreement, shows the relevant sources and their status, preserves chronology, respects permissions and can communicate uncertainty before requesting human confirmation.
Trust is visible behaviour
Trustworthy AI is sometimes described in abstract principles.
Governance gives those principles a practical test.
When the evidence is clear, the system should make the path to that evidence easy to inspect. When the evidence is incomplete or conflicting, the system should make that limitation visible too.
Confidence should follow the record, not outrun it.
Chaird is testing an approach to governance intelligence in which evidence, status, permissions, uncertainty and human authority remain explicit parts of the experience.