Baseline diagnostic findings¶
These are descriptive results from resolver-v1, not accuracy estimates. The
source of truth is data/audit/evaluation.csv.
Four-state aggregate tables¶
State |
Weighted records |
Selected position |
Coverage |
Single-token share |
Selected candidate in relative |
|---|---|---|---|---|---|
Bihar |
69,943,602 |
Final |
97.42% |
2.58% |
35.13% |
Rajasthan |
42,680,004 |
Final |
32.86% |
67.14% |
7.54% |
Maharashtra |
83,029,252 |
First |
98.27% |
1.73% |
89.20% |
Punjab |
18,748,931 |
Final |
88.89% |
11.08% |
37.57% |
Maharashtra’s first and final candidates differ in 98.12% of weighted records. Among those records, the exact relative-name overlap falls into these cells:
Overlap |
Weighted records |
Share of all weighted records |
|---|---|---|
First only |
22,575,921 |
27.19% |
Final only |
2,546,251 |
3.07% |
Both |
50,936,959 |
61.35% |
Neither |
5,405,719 |
6.51% |
The first-only cell is 8.9 times the final-only cell. Exact matching makes this
comparison conservative for first-position variants: jadhab asha ashok and
jadhav ashok, for example, count as final-only because jadhab does not equal
jadhav. Candidate selection does not repair that spelling; variant
canonicalization is a separate stage. Rajasthan’s low coverage is largely a
tokenization fact in the Romanized aggregate table and motivates the
independent Devanagari linked-record check.
Accepted external links¶
Upnaam materialized 4,387 Bihar land links, 536,536 Rajasthan T1 ration links, and 467,882 Rajasthan T2 ration links. These counts exactly match the accepted upstream artifacts.
Accepted links |
Written final token |
Ration-supported exact suffix split |
Abstained |
|---|---|---|---|
Bihar land exact links |
99.52% |
0.00% |
0.48% |
Rajasthan T1 |
32.61% |
3.84% |
63.55% |
Rajasthan T2 |
30.11% |
37.51% |
32.38% |
The Rajasthan results differ substantially by sex and linkage tier:
Tier and sex |
Written |
Exact suffix split |
Abstained |
|---|---|---|---|
T1 female |
31.54% |
8.92% |
59.54% |
T1 male |
33.11% |
1.54% |
65.35% |
T2 female |
34.16% |
16.90% |
48.94% |
T2 male |
28.36% |
46.49% |
25.15% |
The committed evaluation also reports unknown-sex strata and Bihar land results by sex. These are coverage and segmentation diagnostics, not accuracy estimates.
Rajasthan’s T2 result demonstrates the value of a second transcription for
spacing: examples include मोहनसिंह/मोहन सिंह and
सुनितादेवी/सुनिता देवी. It does not demonstrate omitted family surnames,
because the upstream person link requires equality of the complete name
skeleton. Consequently the current family-surname candidate count is zero.
The earlier complete-link research comparator contains 104,215 eligible substitution operations and 8,259 node-to-representative assignments. Those assignments are no longer the canonicalization output because a global partition suppresses one-to-many ambiguity.
The surname-only anchored pilot starts from all 1,004,418 accepted Rajasthan T1/T2 links. Both sides have a final eligible token for 257,509 links, including 230,716 exact token pairs. After aggregation, 2,021 observed roll forms have 2,771 anchor candidates. At the recorded support threshold of 2 and normalized similarity threshold of 0.75, the decisions are:
Decision |
Observed forms |
|---|---|
Accepted identity |
608 |
Accepted variant |
93 |
Ambiguous |
114 |
Unresolved |
1,206 |
These are operating outcomes, not precision estimates. In particular, the accepted links are selected on complete-name skeleton equality, and the ration-side token is a provisional anchor rather than a hand-adjudicated gold spelling. The 114 ambiguous forms demonstrate why candidate rankings are kept separate from decisions.
Bihar official-land reference labels¶
The 4,387 accepted Shekhpura links produce 4,366 official-land reference surnames under the Bihar final-token rule. Of these, 4,365 agree with the roll-side normalized final token. One official land record adds a reference where the roll rule abstains, 20 land names themselves yield a positional abstention, and one token-order conflict is excluded. The accepted labels cover 78 distinct normalized reference surnames.
This near-perfect agreement is expected because the upstream person link used the complete normalized name and relative name. It supports the quality of the official land transcription and the positional extraction on the linked slice; it is not an independent spelling-variant accuracy estimate. A name-withheld relative-and-location linkage was tested and rejected after only 10.7% of 2,696 nominally unique pairs agreed on the held-out full name.
Rajasthan ration-card reference labels¶
Treating the accepted ration transcription as provisional_gold produces
474,587 reference labels from 1,004,418 T1/T2 person-link rows. Another 528,672
rows abstain because the ration name has only one eligible token. The input has
570 ration members linked to more than one elector; all 1,159 associated rows
are retained but excluded.
Among accepted labels, 230,503 equal the roll-side normalized final token,
26,751 differ, and 217,333 supply a token where the roll positional rule
abstains. These are consequences of the declared final-token rule, not proof
that every selected token is hereditary: राम, लाल, देवी, बाई, and
कुमार are common outputs and can play other roles in a name.