Human Review Model¶
Status: Current shared contract
Authority: Current contract and table-review implementation behavior
Last reviewed: 2026-08-14
Question¶
How does human review change Drawing Lab evidence without rewriting history?
Short answer¶
Human review is append-only evidence. A reviewer may accept, adjust, reject, or add a missing region. Software tests verify behavior. Software tests do not validate drawing accuracy.
Review actions¶
| Action | Meaning |
|---|---|
| Accept | The proposal is useful as reviewed |
| Adjust | The reviewer saves a corrected boundary as new accepted truth |
| Reject | The proposal is not useful |
| Add missing | The reviewer adds a region the automatic path missed |
| Unsure | The reviewer records uncertainty without forced acceptance |
Why append-only matters¶
Append-only review keeps earlier proposals visible. A corrected boundary does not erase the proposal history. Corpus baselines can count accepted tables, adjusted boundaries, rejected proposals, and manually added missing tables. Those counts are practical review proxies. They are not formal precision or recall.
Stage boundary¶
Review can correct Detection evidence. Review does not by itself finish Understand assignments. Validate still needs representative reviewed truth before accuracy claims.
Open Questions¶
- Which review fields become mandatory before a sheet counts toward the
30-50sheet fine-tuning gate?
Sources¶
- Current contract:
table-dimension-detection-architecture.md(2026-08-12) - Table review behavior described in the Detection worktree review rail
- Table Review and Later Understanding