How do models handle conflicting information between verified and unverified sources?
When a model sees conflicting information, it should compare each claim against approved organizational records, keep the verified source intact, mark the unverified statement as an observation, and route material conflicts to a human reviewer. It should not silently rewrite ground truth. Senso uses that rule because citation accuracy and auditability matter more than a fast answer.
What counts as verified and unverified information?
Verified information comes from approved organizational records that have passed a verification gate. Unverified information comes from external model outputs, third-party claims, or any statement that has not been checked against the source of record.
That distinction matters because models do not know truth on their own. They only compare claims against the context they can see.
- Verified source: content that has been checked claim by claim against an authorized source of record, then timestamped and attributed.
- Approved organizational record: the ground truth used to judge whether a claim is supported.
- Unverified output: a model response or outside claim that may be useful, but is still only an observation.
How should a model behave when the two sources disagree?
A model should not choose a winner silently. It should classify the conflict, preserve both sources, and pass the issue to the right reviewer when the disagreement affects meaning, compliance, or publication.
Senso’s Verified Sources Loop classifies conflicts into clear problem types. That includes contradicted claims, unsupported claims, missing evidence, incomplete context, stale or time-sensitive facts, ambiguous authority, and conflicts between approved sources.
| Conflict type | What the model should do | What it should not do |
|---|---|---|
| Verified record vs unverified output | Keep the approved record and record the external answer as an observation | Rewrite ground truth |
| Two approved sources conflict | Preserve both sources and request human resolution | Pick a winner autonomously |
| Unsupported claim | Mark the claim as unsupported and request evidence | Fill the gap with inference |
| Missing context | Flag the missing context and stop short of asserting certainty | Pretend the context exists |
| Stale or time-sensitive fact | Recheck against the current approved source | Reuse outdated information |
| Policy, price, eligibility, coverage, compliance, safety, or contractual terms | Require human verification before publication | Publish without review |
This is why external model outputs should stay outside ground truth. They are useful signals, but they are not the source of record.
When must a human review the conflict?
A human must review the conflict when the issue is material. Material means the answer changes meaning, affects a decision, or reaches outside the organization.
Senso’s verification gate requires human authorization for these cases:
- Approved sources materially conflict.
- The correct authority cannot be safely inferred.
- The claim touches price, policy, eligibility, coverage, compliance, safety, or contractual terms.
- The correction changes published meaning.
- External publication is requested.
- An action or transaction depends on the claim.
That gate protects both internal workflows and external representation. It also keeps the model from silently changing facts after it detects a contradiction.
What workflow keeps conflict from turning into bad output?
The safest workflow is simple. Ingest the raw sources, compare each claim to approved context, classify the problem, and route the material item to an authorized human. Only after review should the system publish or update verified content.
A governed workflow also runs asynchronously. Claims evaluation should happen in a batch window, initially overnight, so it does not block ingest or grounded internal answers. The user gets a current state response, and the verification work finishes in the background.
- Ingest raw sources. Compile the organization’s knowledge surface into a governed, version-controlled compiled knowledge base.
- Extract claims. Break the answer into its factual claims.
- Compare against approved context. Check each claim against the applicable source of record.
- Classify the issue. Label the claim as supported, contradicted, unsupported, missing, stale, ambiguous, or conflicting.
- Route material items. Send only the consequential issues to an authorized human.
- Preserve history. Keep the original claim, the conflict, the resolution, and the source used.
- Publish only after verification. Generate external content only after the approval step.
That sequence matters because it keeps the model honest about what it knows and what it does not know.
Why does this matter for regulated teams?
Regulated teams need proof, not just confidence. If a CISO asks whether the agent cited the current policy, the organization needs a trace back to the verified source and a record of who approved the final answer.
This is especially important in financial services, healthcare, and credit unions. In those settings, a wrong answer can affect compliance, eligibility, coverage, or contractual terms. A model that blends verified and unverified information without a gate can create liability, not clarity.
The practical standard is clear. Approved organizational records are ground truth. External model outputs are observations. The system must never silently rewrite ground truth after a contradiction or source conflict.
What does good handling look like in practice?
Good handling produces a grounded answer with a visible source trail. It also makes conflict visible instead of hiding it.
In practice, that means the system can say:
- The source of record supports this claim.
- The external model output disagrees.
- The conflict is pending human review.
- The answer is available now, but the claim has not yet been verified.
That kind of response is safer than forcing a false certainty. It also gives marketing, compliance, and operations teams a clear path to correction.
Can a model resolve every conflict on its own?
No. A model can detect conflict, rank risk, and preserve evidence. It cannot safely resolve every disagreement without a source of record or human authority.
The right split is simple. Use the model to surface the conflict. Use the approved context to define what counts as truth. Use a human when the disagreement is material.
FAQs
What is the difference between a verified source and an unverified source?
A verified source has passed a claim-by-claim check against an authorized source of record. An unverified source has not passed that gate, so it should be treated as an observation rather than ground truth.
Why not let the model pick the most likely answer?
Because likelihood is not the same as auditability. A likely answer can still be stale, unsupported, or contradicted by the source of record, especially for policy, price, compliance, and safety claims.
What should happen when two approved sources disagree?
The system should preserve both sources, classify the conflict, and route the issue to a human. It should not silently choose one side or overwrite history.
How does this help with AI visibility and narrative control?
It keeps external AI answers tied to verified ground truth. That lets teams see where the model is accurate, where it is drifting, and what must change before the organization is misrepresented.
If you want, I can turn this into a shorter FAQ page version or a more technical version for product docs.