v0.1.0 PyTorch checkpoint × golden v0.1.2 (apples-to-apples baseline)

• entries evaluated: 4535
• full-parse exact match: 0.0088
• mean token confidence: 0.8587

Per-component F1

tag	precision	recall	f1	support
country	0.0135	0.1388	0.0246	245
region	0.0333	0.0409	0.0367	3205
locality	0.0677	0.0909	0.0776	3357
dependent_locality	0.0000	0.0000	0.0000	40
postcode	1.0000	0.0003	0.0007	2980
subregion	0.0000	0.0000	0.0000	0
cedex	0.0000	0.0000	0.0000	1

Calibration (confidence bucket → accuracy)

bucket	n	accuracy
0.0–0.1	0	0.0000
0.1–0.2	1	1.0000
0.2–0.3	208	0.0913
0.3–0.4	1660	0.1084
0.4–0.5	3207	0.1366
0.5–0.6	4030	0.1846
0.6–0.7	3981	0.2042
0.7–0.8	4206	0.2183
0.8–0.9	5627	0.2472
0.9–1.0	38328	0.3366