MirmiComparability criteria
Commercial context and risk can make two textually similar clauses materially different. Search should combine content, metadata and approval history.
- document type and clause;
- entity, counterparty and segment;
- date, jurisdiction and applicable policy;
- decision and exception status.
What the answer should show
Every result should lead back to the document, passage and approval. A summary speeds up reading but is not enough to support the decision without evidence.
Reuse without accidentally creating a rule
The workflow should distinguish a qualified precedent, a one-off exception and an approved position. An exception signed in a specific context should not automatically raise future acceptance thresholds.
A false-positive example
Two agreements contain the same liability cap. One covers a three-month, low-value pilot; the other a multi-year critical service involving sensitive data. Textual similarity is high, but decision comparability is low. A useful result exposes that difference before presenting either matter as a precedent.
Conversely, different wording can encode the same commercial compromise. Retrieval therefore combines text, metadata and approval history rather than relying on semantic ranking alone.
Qualify and measure the workflow
For each precedent, preserve at least the source, passage, context, decision owner, date, outcome and status: standard position, qualified precedent or exception. Lawyers should be able to correct comparability, and those corrections should improve later retrieval.
- time needed to find and verify a precedent;
- share of results opened back to the source passage;
- Legal correction or rejection rate;
- share of decisions recorded with complete context.