Bounding repair scope after Jev flags an extracted field

Goal

Keep an extraction cascade from rewriting an entire record when only a specific field lacks support, and preserve the distinction between missing evidence and an incorrect value.

Prerequisites

Have source material, an extraction schema, field-level evidence references, and an explicit escalation policy. The repair service must have authorized access to the source; a classifier result alone is not evidence.

Steps

1. Validate record structure before semantic review. List missing, malformed, and unsupported fields separately so a repair request does not obscure the original kind of failure.

2. Use narrow verification questions about each relevant field, following the vendor cascade pattern. Ask whether evidence supports the supplied value and whether an empty value omitted information actually present.

3. Construct a repair packet limited to the affected fields and their necessary context. Include constraints between fields where they genuinely interact, while protecting already supported values from incidental rewriting.

4. Revalidate the repaired record against the source and the schema. Preserve the original extraction identifier, the field changes, and the reason for escalation without retaining unnecessary private text in logs.

5. Exercise absent information, contradictory passages, template examples mistaken for facts, and repairs that alter previously correct fields. A second model response must not count as independent truth merely because it sounds more confident.

Expected result

The final record distinguishes accepted fields, repaired fields, and unresolved fields. The consumer can see whether the cascade obtained new evidence or merely produced another interpretation of the same source.

Limits and test basis

This is a proposed workflow. The vendor cookbook reports its own cascade experiment, which is not reproduced here. Neither a semantic verifier nor a more expensive generator guarantees factual correctness or lower cost for another workload. The underlying interface or pattern is described in SDE cascade; the workflow above is a proposed adaptation.