[WSJF-8] Parser quality self-test: verify new parsers against source email #3

Open
opened 2026-06-27 10:45:03 +02:00 by bullitt · 0 comments
Owner

Context

Self-evaluation bias: AI 'confidently praises' its own field_maps even when quality is mediocre.

Problem

AI generates a field_map and we store it without verification. The parser might not actually work on the email that created it.

Proposal

  • After AI generates field_map, immediately apply it to the same email body
  • Compare deterministic extraction result with AI's direct extraction
  • If they disagree on >1 field, mark parser as low-confidence or discard it
  • Log mismatches for debugging

WSJF: 8

## Context Self-evaluation bias: AI 'confidently praises' its own field_maps even when quality is mediocre. ## Problem AI generates a field_map and we store it without verification. The parser might not actually work on the email that created it. ## Proposal - After AI generates field_map, immediately apply it to the same email body - Compare deterministic extraction result with AI's direct extraction - If they disagree on >1 field, mark parser as low-confidence or discard it - Log mismatches for debugging ## WSJF: 8
Sign in to join this conversation.
No labels
arch
harness
security
ux
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
bullitt/trackbox#3
No description provided.