Testing results¶
T1 and T3 are completed feasibility tests. T2, T4, and T5 results will be added as re-runs complete.
T1 — Glass schedule with auxiliary table¶
T1 tests the full specialist pipeline against a multi-row glass schedule that contains a glazing auxiliary table, compound references, subtype rows, and item card diagrams all on a single page. The Park Lane document is a representative test case because it exercises all four specialist strategies simultaneously and includes both straightforward rows and complex compound-reference resolution.
Document: CRPA-Park Lane T1.pdf
Page tested: 1
Template: Glass Schedule
Columns verified: Glass Brand, Glass Type, Glass Arrangement Configuration, Frame Brand, Width, Height, Quantity, Special Notes
Verdict¶
| Check | Result |
|---|---|
| Row count | 19 rows (correct) |
| Subtype inheritance | P1a, P1b, P2b, P3a, P4a, P8a, P12b, P13a inherit base type dimensions and Glass Arrangement Configuration |
| Compound reference resolution | GL-03/GMT-01 correctly resolved; secondary GMT-01 details carried into Special Notes |
| BASIS OF DESIGN PRODUCT exclusion | Absent from all Special Notes |
| IBC notes application | Applied consistently, no duplication |
| P5/P6 exclusion | Correctly excluded — both marked NOT USED in diagrams |
| Strategies fired | auxiliary_table, text_rule, image_legend, dimension_card |
Note
The debug artifact for this run is debug/run_20260330_143028.json. This file is the final T1 test artifact.
T2 — General notes and text rules¶
Coming soon
T2 results will be added after re-runs are complete.
T3 — Image legend enrichment¶
T3 tests the image_legend specialist's ability to match style codes from door and window legend diagrams and translate them to exact operability enum values. VTT Fairway is the primary T3 test document — it has a six-style door legend and a three-style window legend on a single page, with eight sheet notes that must be scoped correctly.
Document: VTT - Fairway T3.pdf
Page tested: 1
Template: Standard Takeoff
Strategies fired: text_rule, image_legend
Verdict¶
| Check | Result |
|---|---|
| Row count | 21 rows (correct) |
| Operability enum strings | All exact enum values, no raw abbreviations |
| D14 Sectional / Overhead | Correct — Sectional / Overhead enum resolves cleanly |
| Text rule scoping (Sheet Notes 6/7) | Note 6 applied to exterior doors only, Note 7 to all doors |
| W4 tempered glass rule | Sheet Note 1 applied correctly |
| D11 operability | Returns Swing Double — ground truth is Sliding Door |
| D13 operability | Returns Swing Double — ground truth is Sliding Door |
| Product Type — doors | Empty for most door rows — porting-level gap, not a Cartex issue |
D11 and D13 are both Style D doors — a sliding patio door configuration. The pipeline returns Swing Double for both. The root cause is that the extraction prompt describes Style D as a "double door with full-length glass panels" without identifying the sliding track visible in the diagram. The enrichment specialist then maps this to Swing Double — the closest textual match — rather than Sliding Door. This is a known limitation of the current image interpretation prompt for sliding door visuals. The fix is deferred and documented in the limitations section.
Warning
D11 and D13 operability is incorrect in this run. Ground truth is Sliding Door. This miss is tracked as a known limitation.
Note
The debug artifact for this run is debug/run_20260330_162655.json.
T4 — Dimension card enrichment¶
Coming soon
T4 results will be added after re-runs are complete.
T5 — Multi-page extraction¶
Coming soon
T5 results will be added after re-runs are complete.