Late source changes are ordinary in high-stakes work. The operational failure is not that a document changed. It is that the team cannot see which downstream work depended on the old version.

A revised SPA replaces the prior draft. Five clauses move. Two findings quote language that no longer exists. One memo section still uses the old indemnity cap. A redline is waiting for counsel, and a partner has already opened yesterday’s export.

A search index can ingest the new file. That is necessary, but it does not close the loop.

Start with source identity, not text similarity

The system needs a stable identity for the business object and a distinct identity for each version. SPA_Final_v2.pdf and SPA_Final_v3.pdf may be two files, but the team understands them as versions of one agreement.

Useful metadata includes:

  • provider and external identifier;
  • workspace and folder location;
  • uploader or sync actor;
  • received and processed timestamps;
  • file hash and document family;
  • superseded and superseding relationships;
  • extraction status and processing errors.

Text similarity can help propose a relationship. A reviewer or deterministic rule should confirm consequential version links.

Build lineage at the point of use

Lineage becomes valuable when evidence is attached to work. A citation should identify not just the document but the precise location used by the claim. Numeric analysis should retain workbook, sheet, range, and run context. A meeting observation should retain transcript position and speaker.

The next links are equally important:

Evidence to claim. Which finding, matrix cell, clause, answer, or model input relies on this evidence?

Claim to review. Who accepted, rejected, edited, or overrode it, and under which policy?

Claim to output. Which memo section, redline, client update, or export includes it?

Without these links, change detection produces an alert about a file. With them, it produces a review queue the team can act on.

Stale does not mean wrong

Staleness should be a state that asks for attention, not an automatic factual judgment.

The new agreement may leave the cited clause unchanged. A refreshed workbook may alter formatting but not values. A new management comment may add context without contradicting the prior finding.

A practical state model is:

  1. Current: all supporting evidence is current.
  2. Potentially stale: a supporting source changed or was superseded.
  3. Under review: an owner is comparing the old and new evidence.
  4. Reaffirmed: the conclusion remains valid against the new source.
  5. Revised: the claim changed and dependent work must update.
  6. Retired: the claim should no longer be used.

This vocabulary separates a system signal from a professional conclusion.

Route impact by materiality

Not every change deserves the same interruption. A punctuation change in a low-risk exhibit should not page the engagement lead. A revenue restatement or new change-of-control clause might block export.

Routing can consider:

  • the materiality of the dependent claim;
  • whether the claim appears in an external deliverable;
  • the stage of the deal or matter;
  • the reviewer role required;
  • time remaining to a milestone;
  • whether other evidence still supports the conclusion;
  • whether the change created a conflict.

The output is not a bigger notification feed. It is a prioritized set of review tasks with enough context to make a decision.

Test with a controlled change event

Many pilots benchmark answer quality against a frozen source set. That misses the operating problem.

Add a second test:

  1. Load a representative set of sources.
  2. Produce findings and a draft deliverable.
  3. Replace one important source with a realistic revision.
  4. Measure whether the system identifies affected evidence and claims.
  5. Verify that the right reviewer receives a useful task.
  6. Confirm that blocked outputs cannot silently ship.
  7. Reaffirm or revise the claim and inspect the retained record.

Track false positives as well as missed impact. A system that marks everything stale will be ignored. A system that only catches exact quotation changes will miss semantic impact.

The operating principle

Professional work is temporal. A claim can be well supported on Monday and unsafe on Thursday without anyone making a mistake.

The system should therefore preserve not only where a conclusion came from, but which version of the evidence supported it and what happened when that evidence changed.

That turns a late upload from a scavenger hunt into a governed review event. It also gives the final output a more honest status: ready based on the evidence the team has actually reviewed, not merely the evidence the platform has most recently indexed.

How we approach this topic

This field note is based on the workflow and product-design questions we encounter while building Underlying. It is educational, not legal, investment, or security advice. Product examples describe design patterns unless explicitly stated as generally available.

Reviewed by Underlying Product Team.