Validate¶
Status: Validation gates are written; representative accuracy remains unvalidated
Authority: Current contract validation gates
Last reviewed: 2026-08-14
Question¶
What does Validate receive, do, produce, and refuse to claim?
Short answer¶
Validate measures Drawing Lab results against human-reviewed truth. Software tests verify behavior. Software tests do not validate drawing accuracy.
Receives¶
- Understood objects and incomplete evidence that remains visible
- Append-only human review records
- Evaluation metrics from representative sheets
- Owner acceptance decisions
Does¶
- Separates verified behavior from validated accuracy
- Measures table region quality, structure quality, and cell-text quality when truth exists
- Measures dimension component recall, complete-dimension recall, false candidates, attachment accuracy, and parsing accuracy when truth exists
- Records cold and warm runtime separately before speed claims
- Keeps review proxies such as accepted, adjusted, rejected, and manually added counts labelled as proxies
Produces¶
- Validated claims only when representative human-reviewed truth supports them
- Accepted product boundaries after owner decision
- Explicit non-claims when evidence is incomplete
Does not claim¶
- That a preferred provider choice is a production accuracy score
- That local timing on one machine is a product SLA
- That review-rail counts are formal precision or recall
- That Detected Sheet completeness equals accuracy
Current contract gates¶
Tables require region precision and recall, false positives and misses per sheet, boundary quality, structure accuracy, and cell-text accuracy.
Dimensions require component and complete-dimension recall, false candidates per sheet, feature-attachment accuracy, parsing accuracy, and controlled cold and warm runtime.
Fine-tuning stays deferred until roughly 30-50 representative sheets are fully reviewed and failure families justify it.
Next stage receives¶
Reviewed evidence and truth feed product claims, later experiments, and accepted decisions.
Open Questions¶
- Which first sheet corpus will open formal table and dimension validation?
Sources¶
- Current contract:
table-dimension-detection-architecture.md(2026-08-12), Validation gates - Status vocabulary: Status Vocabulary