The source-of-truth rule reduces AI hallucinations by forcing every material claim to resolve to an authoritative, current source before it is presented as fact. The model may summarize or explain the evidence, but it may not invent authority, average conflicting records, or turn missing evidence into a positive statement. When the source cannot support the claim, the correct output is qualification, escalation, or refusal.
Assign authority to claims, not file formats
Start with claim types. Price may belong to a product catalogue, customer status to CRM, contractual obligation to the signed agreement, and an operating decision to an approved log. A polished PDF or frequently shared spreadsheet does not become authoritative unless the accountable owner has designated it for that purpose.
Create a source registry that maps claim type and entity to the canonical system, owner, effective-date behavior, sensitivity, and conflict route. Include exact identifiers and aliases so the system does not confuse similar customers, products, or legal entities. The registry should be deterministic enough to review outside the model.
Derived indexes keep references to the registry and source. They can rank likely material, but the selected evidence is hydrated from the current record and checked against the expected version. The architecture described in company memory beyond vector search gives public context for this pattern without proving outcomes in another deployment.
Authority also has scope. A current regional price list may be correct for one market and wrong for another. A signed contract may govern one client while the template has only drafting authority. Source selection must combine claim type, entity, purpose, and time.
Represent conflict and absence honestly
Consider a current price list that conflicts with an older proposal template. The system should not ask the model which wording sounds more confident. It should recognize different authority levels, cite the current list for the price, and flag the template for correction if its stale value can mislead future work.
If two eligible authoritative sources disagree, return a conflict state with both references, effective dates, and the named owner who can decide. Do not merge values or hide disagreement behind a vague sentence. Until resolution, high-consequence use may need to stop.
A claim record can contain the normalized claim, source identifier, passage or field, source hash, retrieval time, entity, policy decision, and output status. Status distinguishes directly supported, inferred within a rule, conflicted, missing, and prohibited. This makes refusal a designed result rather than an unexpected model failure.
The verified completion memory release documents public work on durable evidence and final-state checks. It does not certify the truth of every source or provide a universal hallucination rate.
Test whether unsupported claims fail closed
Build cases around real claim risks: a stale but popular file, similar entity names, a missing field, two current records in conflict, a source outside the user’s permission, and an inferred statement that goes beyond the passage. Define the accepted source and allowed wording before the run.
Measure material answers with valid provenance, conflicts detected, unsupported positives blocked, outdated references rejected, and corrections after source change. Review severe errors separately from style. A response can be concise and still pass if it accurately says the evidence is unavailable.
Business-context retrieval described in the Unlimited Skills retrieval release is public evidence of maintained tooling. The product still needs its own registry, authorization, dataset, and independent reviewer.
After an owner resolves a conflict, update the canonical source or registry, invalidate affected index entries, and rerun regression cases. Do not patch only the prompt. The repair belongs where authority and evidence are controlled.
Frequently Asked Questions
The domain owner designates the authoritative system for that claim type and resolves conflicts that the retrieval layer cannot settle.
Record the source's authority rank, version, effective date, conflict status, and citation. Together they show why one record supported the claim while another was rejected or escalated.
It should return both references, their effective dates, and the named owner of the decision, then stop consequential use until the conflict is resolved.
Record the normalized claim, source and version, supporting field or passage, retrieval time, entity, policy decision, and a status such as supported, conflicted, missing, or prohibited.
AI4SALE has built source-backed company memory and bounded retrieval mechanisms and documents their contracts, without claiming perfect factuality or customer outcomes. Teams can use the exact planned AI4SALE enterprise AI search service after its publication gate to define source rules, conflict handling, and fail-closed tests.