Document-Revision Conflicts with NLI

Objective

A summary repeats "retention is 30 days" from revision 1; revision 2 changed it to 90. This recipe shows the two-step pattern that works — and why numeric, temporal claims are the case where you must not lean on NLI alone.

Step 1: Both revisions as anchors — the diagnostic

curl -s -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{
    "model": "nli-default",
    "statement": "The archive retention period is exactly 30 days.",
    "truth_anchors": [
      "Revision 1 states that the archive retention period is exactly 30 days.",
      "Revision 2 states that the archive retention period is exactly 90 days."
    ],
    "closed_world": false,
    "contradiction_threshold": 0.6, "entailment_threshold": 0.6
  }' http://localhost:8080/v1/ai/nli/verify

Observed on v2.1.1-ce with nli-default: entailment 0.98 on revision 1, contradiction 1.00 on revision 2 → verdict LIE (contradiction takes priority). The revisions pull in opposite directions, and the conflict is flagged. Numeric wording is where NLI models miss most often, so treat this as a conflict detector and record what your model actually returns on your own wording before automating anything on it.

Step 2: Select the active revision first — the production pattern

curl -s -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{
    "model": "nli-default",
    "statement": "The archive retention period is exactly 90 days.",
    "truth_anchors": ["The active archive policy sets the retention period to exactly 90 days."],
    "closed_world": false,
    "contradiction_threshold": 0.6, "entailment_threshold": 0.6
  }' http://localhost:8080/v1/ai/nli/verify

Expected (verified): verdict TRUTH, entailment 0.98. Your application selects which revision is active — the classifier does not know that later documents supersede earlier ones.

Honest boundaries

For an authoritative retention value, parse the active revision into a typed number and compare with SQL — WHERE retention_days = 90 cannot be sweet-talked by phrasing. Use NLI for the wording around the number, not the number.

Key Concepts Learned

• Revision selection is application logic; NLI sees only the text supplied.
• Numeric and temporal claims deserve typed comparison; NLI is the second-opinion layer.