Saphan
Guides & FAQ

KNOWLEDGE / UPDATING

Cases evolve.
The answer must date its sources.

An answer that was correct yesterday may be incomplete today. Here, temporal intelligence means organizing the review of a case when new documents arrive, with a record of what has changed and what remains uncertain.

Written by Saphan · 7 September 2026

Re-examine a case without changing it on its own

The workflow described here is a line of work to explore, not a Saphan capability already demonstrated. It assumes neither a model that learns on its own nor an improvement guaranteed overnight.

The envisaged mechanism is explicit: select the eligible cases, re-read the new documents, prepare a difference and submit the changes to the planned review. The processing can be scheduled or triggered by an event; its frequency should serve the business.

Distinguish the dates that change the decision

The date a document is received, the date it is modified and the date it applies are different. A document received today may correct an old event or announce a future rule. A summary should state “known as of this date” and “applicable to this period” when the distinction matters.

Fictional example: a supplier sends a new delivery time that applies to upcoming orders. The reprocessing flags the quotes concerned without changing commitments already accepted. A person checks the cases actually affected.

Produce a difference before changing the reference

Each run keeps the identifiers and versions of the documents read, the version of the processing, the previous result and the new proposal. A contradiction stays visible: the system must not pick a source only because it is more recent.

The proposal goes through a review status. A case already processed must not receive the same action twice. Source deletions and access revocations must also trigger a suitable update.

Test a complete cycle with Saphan

Pick a small batch of cases and one stable question. Introduce a new document, a retroactive correction and a missing source. Check the differences proposed, the actions avoided and the review workload. The first result to expect is a review that can be understood and replayed, not a promise of growing intelligence.