Condition: Guided — models given the questions plus expert domain guidance
Figures shown: mean accuracy across all 10 runs, per page

Page generated: 2026-08-16
Data (model runs) generated: 2026-04-11
Scoring inputs: ground truth cc2ed8f5, evaluate.py 845b70e9

Image Complexity vs Model Accuracy

Two ways to measure page complexity: structural (counts derived from ground truth) and empirical (mean model accuracy). Pages with denser content are systematically harder.

Correlation: complexity vs mean accuracy

Pearson r across 39 pages. Negative = denser/more complex pages have lower accuracy. r² gives the fraction of accuracy variance explained.

Structural complexity (from ground truth)

Complexity measure	Range	Pearson r	r²
Text density (travelers + con + e)	7–118	−0.751	0.563
Named travelers (Q13)	4–62	−0.773	0.598
Unnamed travelers (Q14)	1–39	−0.611	0.373
Brackets (Q19)	0–13	−0.558	0.311
con instances (Q9)	1–16	−0.664	0.442
e conjunction instances (Q11)	0–8	−0.590	0.348
Places of origin (Q15)	2–25	−0.774	0.599
Places of stay (Q16)	2–13	−0.482	0.233

Visual complexity (from images via Sauvola binarization)

Complexity measure	Range	Pearson r	r²
Ink density (Sauvola foreground ratio)	0.042–0.137	−0.714	0.510
Connected ink components (≥5px)	423–2695	−0.771	0.595
Median component area (px²)	12.000–84.000	−0.261	0.068
Mean component area (px²)	49.800–153.020	+0.331	0.110
Background entropy (staining/bleed proxy)	2.811–3.847	−0.653	0.427
Histogram mid-weight (bleed proxy)	0.169–0.966	−0.033	0.001

Visual measurements: Sauvola binarization (window=25, k=0.2) computed on downscaled images (1500px max). Ink density = foreground pixel ratio. Components = connected ink regions ≥ 5px. Background entropy ≈ local intensity variance over non-ink pixels (proxy for staining). Histogram mid-weight = fraction of pixel-intensity histogram in the [0.25, 0.75] band (proxy for bleed-through; clean documents are bimodal so this is small).

Scatter: complexity vs accuracy, by model

Each point = one page for one model. Click legend entries to toggle models.

Text density (structural)

Connected ink components (visual)

Per-page complexity ranking

Click any column header to sort.

Page▲	Mean accuracy▲	Text density▲	Named▲	Unnamed▲	Brackets▲	con▲	e▲	Ink density▲	Components▲	Bg entropy▲	Hist mid▲
11	0.502	118	62	33	11	15	8	0.0905	2633	3.555	0.251
37	0.483	108	49	35	13	16	8	0.0847	2502	3.791	0.241
8	0.487	99	43	39	6	14	3	0.0994	2691	3.571	0.232
10	0.508	97	53	26	10	15	3	0.0891	2480	3.571	0.235
36	0.540	73	37	22	8	10	4	0.0961	2585	3.734	0.249
12	0.601	72	38	18	6	9	7	0.0760	1710	3.131	0.169
27	0.494	67	35	17	7	12	3	0.0824	2629	3.601	0.209
2	0.464	65	35	18	4	9	3	0.1063	2695	3.626	0.201
1	0.569	56	34	12	5	4	6	0.0928	2136	3.815	0.966
34	0.513	56	29	16	6	8	3	0.0772	1778	3.769	0.236
19	0.533	53	26	15	3	7	5	0.0804	2308	3.510	0.241
9	0.619	51	33	11	2	3	4	0.0559	1614	3.619	0.278
22	0.534	51	21	16	2	11	3	0.0844	2366	3.407	0.249
23	0.482	49	33	8	6	6	2	0.1062	1989	3.699	0.298
21	0.595	48	27	12	6	4	5	0.1209	1857	3.637	0.412
4	0.506	46	28	8	0	7	3	0.1374	1596	3.580	0.377
6	0.508	46	26	10	4	8	2	0.0917	2597	3.434	0.263
17	0.677	43	14	23	1	5	1	0.0770	1530	3.519	0.195
20	0.656	39	22	10	5	7	0	0.0730	2157	3.215	0.198
14	0.644	38	22	8	0	5	3	0.0884	1838	3.485	0.258
38	0.606	38	31	3	7	1	3	0.0800	1915	3.385	0.197
26	0.570	36	24	6	1	2	4	0.0930	1818	3.835	0.236
28	0.558	36	19	11	1	4	2	0.0596	1614	3.847	0.217
16	0.515	35	20	10	3	5	0	0.0715	1935	3.368	0.216
24	0.524	32	22	3	4	1	6	0.0914	1656	3.426	0.221
32	0.696	31	16	7	2	7	1	0.0787	1800	3.357	0.248
33	0.624	28	18	5	3	4	1	0.0628	1939	3.486	0.219
25	0.726	27	12	9	2	6	0	0.0569	1102	3.751	0.187
29	0.807	25	13	6	0	2	4	0.0535	1438	3.225	0.197
30	0.745	24	22	1	3	1	0	0.0577	1305	3.439	0.200
5	0.628	20	11	4	0	2	3	0.0858	1218	3.608	0.440
18	0.822	20	11	5	2	3	1	0.0685	1225	3.110	0.224
31	0.776	19	14	3	4	2	0	0.0602	1897	3.390	0.223
35	0.694	19	17	1	2	1	0	0.0740	1571	3.637	0.184
39	0.874	15	10	3	1	2	0	0.0578	1706	3.090	0.269
3	0.854	11	4	5	1	1	1	0.0425	423	3.373	0.416
7	0.864	10	6	2	2	2	0	0.0586	805	3.131	0.362
13	0.819	10	6	3	1	1	0	0.0508	782	2.811	0.207
15	0.835	7	5	1	0	1	0	0.0641	1431	3.176	0.220