Skip to content

Provenance, Cache, and Timing

Status: Current shared contract

Authority: Current contract (2026-08-12)

Last reviewed: 2026-08-14

Question

What must every Drawing Lab result carry so a reviewer can trust reuse and timing?

Short answer

Every result keeps method, settings, runtime, cache, and source provenance. Switching a review lens does not rerun processing. A page result should assemble compatible cached input results when possible.

Provenance fields

  • Method or provider identity
  • Settings and thresholds
  • Source page and Extraction capture identity
  • Review state used during the run
  • Cache identity for reused evidence
  • Stage runtime when the lab records it

Cache rules

Canonical Detection input sets are cached by page, extraction capture, review state, and pagination. OCR and arrowhead caches may warm without invoking the generic detected-sheet assembler. Fast Dimension Lines reuses one page-wide straight-line pass when the cache is valid. Providers and geometry sources are never automatically fused.

Timing rules

Measure cold preparation and warm reuse separately. Measure line extraction, arrow search, association, and full runtime separately before a redesign claims speed. Local timing on Apple M1 Max is evaluation evidence only.

Example evaluation timing

Saved local comparison timing on Apple M1 Max reports:

  • PP-DocLayoutV3 inference median approximately 1.27 s
  • complete proposal plus verification median approximately 1.42-1.48 s per prepared sheet

These numbers are evaluation evidence. They are not a validated product SLA.

Open Questions

  • Which timing stages become the permanent benchmark dashboard for dimensions?

Sources

  • Current contract: table-dimension-detection-architecture.md (2026-08-12)
  • Status vocabulary: Status Vocabulary