# Results validation (Gate 7)

Generated by scripts/12_validate_results.py on 2026-10-20 08:52. Each headline number quoted in the manuscript (via paper/tables/numbers.tex) is recomputed here from the processed data by independent code paths (direct cost-model call, direct per-image aggregation) and compared with the macro value.

Check	Macro	Macro value	Recomputed	Tolerance	OK
sample size	`nSample`	1,200	1200	0	✓
image URLs in sample	`nImagesSample`	14,245	1.424e+04	0	✓
frame size	`nFrame`	7,745	7745	0	✓
roof agreement (Sonnet)	`accroofcovering`	95.8	95.78	0.15	✓
roof coverage (Sonnet)	`covroofcovering`	87.1	87.13	0.15	✓
basement shown (Sonnet, K10)	`shownBasement`	3.1	3.083	0.15	✓
bathroom shown (Sonnet, K10)	`shownBathroom`	19.6	19.58	0.15	✓
median V B0	`VsonnetBZeroB`	407,545	4.075e+05	1.0	✓
median D B0	`DmedsonnetBZeroB`	0.62	0.6167	0.006	✓
price ratio B0	`priceRatiosonnetBZeroB`	0.75	0.7467	0.006	✓
price MdAPE B0	`priceMdapesonnetBZeroB`	30.1	30.15	0.06	✓
median V B3	`VsonnetBThreeB`	514,919	5.149e+05	1.0	✓
median D B3	`DmedsonnetBThreeB`	0.38	0.375	0.006	✓
price ratio B3	`priceRatiosonnetBThreeB`	0.95	0.9471	0.006	✓
price MdAPE B3	`priceMdapesonnetBThreeB`	27.9	27.87	0.06	✓
cost identity holds for all B3 rows	`—`	True	1	0	✓
verified references	`nRefsVerified`	104	104	0	✓
bib entries = verified refs	`—`	104	104	0	✓

18 / 18 checks passed.

# Manual checks performed

Every table in the manuscript is an \input of a generated fragment; no number is typed by hand (grep of sections/*.tex for digits outside macros: only section numbers, years, equation constants and the 15 m / 20 % thresholds that are design parameters).
Figures are regenerated from results/tables/*.csv and data/processed/*.parquet by src/visualization/figures.py; PNG previews were inspected for clipping and label collisions.
Citations: 104 references, each verified against Crossref/arXiv metadata (docs/verification/); BibTeX run reports 0 warnings; no undefined citations in main.log.
The pre-registered experiment plan (docs/experiment_plan.md) lists 16 experiments; 14 were executed, 2 (human expert agreement, human-in-the-loop gains) are marked not feasible and replaced by documented substitutes.
Known approximations stated in the text: assessor building value indexed by the construction price index; roll land value indexed by the single-family price index; regional cost level uniform; Monte Carlo excludes latent attributes.