Bihar land surname counts¶

Estimand¶

The artifact counts the final eligible token written in each distinct Bihar land-record ryot name. One input row is one distinct official full-name string, not a person, account, holding, parcel, or land-record transaction. Repeated people with the same written full name are represented once, while spelling differences remain separate source strings.

The output is a provisional official-record vocabulary. It does not assert that every final token is a hereditary surname or that the most frequent spelling is canonical.

Source and rules¶

The source Parquet contains 3,197,304 rows in name_of_ryot: 3,197,303 unique nonnull strings and one null. The command rejects duplicate nonnull names and non-text name columns.

Apply the Bihar final-eligible-token rule with a two-letter minimum.
Drop only the versioned leading honorifics used throughout Upnaam.
Abstain on null, empty, ineligible, and one-token names.
Retain देवी, रानी, कुमार, सिंह, and similar tokens.
Group with normalization-v1; do not transliterate or edit-merge.
Do not read or use caste annotations.

Run locally with:

upnaam aggregate-bihar-land unique_hindi_names_uncleaned.parquet \
  bihar_land_surname_counts.parquet \
  --audit bihar_land_counts.json --manifest bihar_land_counts.manifest.json

Output dictionary¶

Field	Meaning
`surname_source_normalized`	Conservative same-script grouping key
`surname_raw_mode`	Most frequent exact selected token among distinct full names
`surname_raw_mode_count`	Distinct full names carrying the modal raw token
`raw_variant_count`	Exact raw selected tokens in the normalized group
`distinct_full_name_count`	Distinct official full-name strings selecting the token
`surname_provenance`	`bihar_land_written_final_token`
`normalization_revision`	Normalization implementation revision
`aggregate_revision`	Immutable aggregation-contract revision

Complete source run¶

Quantity	Count
Source rows	3,197,304
Unique nonnull full-name strings	3,197,303
Names with a selected final token	2,920,486
Abstained names	276,818
Normalized final-token groups	144,081

The abstentions comprise 276,318 one-token names, 497 names without an eligible token, and 3 missing or normalization-empty names.

Normalized token	Distinct full-name strings
`देवी`	177,693
`सिंह`	125,117
`यादव`	118,315
`सिह`	73,125
`साह`	69,029
`राय`	65,022
`महतो`	63,231
`मंडल`	58,903
`वगैरह`	41,011
`चौधरी`	34,824

The वगैरह result is a material source-specific failure of the literal final token rule: the land name field commonly appends administrative notation meaning “and others.” Other frequent renderings include वगै0, वोगैरह, वैगरह, वगेरह, and बगैरह; अन्य is another observed suffix. These remain preserved in the written-token baseline. A separate artifact applies the approved source-specific inference rule described below.

Exact terminal-notation inference¶

infer-bihar-land emits a second group-by artifact and never overwrites the literal written-token counts:

upnaam infer-bihar-land unique_hindi_names_uncleaned.parquet \
  bihar_land_inferred_surname_counts.parquet \
  --audit bihar_land_inference.json \
  --manifest bihar_land_inference.manifest.json

Field	Meaning
`surname_inferred_normalized`	Grouping key after the exact suffix rule
`surname_inferred_raw_mode`	Most frequent exact inferred token
`surname_inferred_raw_mode_count`	Distinct names carrying the modal token
`raw_variant_count`	Exact inferred-token renderings in the group
`distinct_full_name_count`	Total distinct source-name support
`written_final_token_count`	Support unchanged from the literal final token
`record_suffix_previous_token_count`	Support moved exactly one eligible token left
`record_suffix_chain_token_count`	Support moved left across an exact connector or repeated suffix
`inference_source`	Frozen official-name-vocabulary source label
`normalization_revision`	Normalization implementation revision
`inference_revision`	Exact suffix-rule revision

The rule is deliberately narrow:

Apply the ordinary Bihar final-token rule.
If the selected final token exactly matches one of अन्य, बगेरह, बगैरह, वगेरह, वगै, वगै0, वगैरह, वगैरा, वैगरह, or वोगैरह, scan left across exact connector tokens एव and एवं and repeated approved suffix tokens, then select the first preceding eligible token in the separate inferred field.
Otherwise retain the written final token as the inferred result.
Do not use edit distance, fuzzy suffix recognition, or another token blacklist. और and तथा are not skipped because the audit found no case where either immediately precedes an approved suffix.

One inferred-token row reports total distinct-full-name support plus separate direct-written, previous-token, and chain-token count columns. Their sum must equal distinct_full_name_count. The artifact also records bihar-land-record-suffix-inference-v2 and the source and normalization revisions.

The complete run adjusts 79,891 names, or 2.74% of the 2,920,486 names with a selected token. Of these, 79,183 use the immediate previous eligible token and 708 use the exact chain rule. It produces 146,835 inferred-token groups. The suffix matches are:

Exact suffix	Adjusted names
`वगैरह`	41,011
`वगै0`	13,723
`वोगैरह`	5,475
`वैगरह`	4,758
`वगेरह`	3,494
`अन्य`	3,500
`बगैरह`	2,410
`वगै`	2,089
`बगेरह`	1,729
`वगैरा`	1,702

The chain audit records 699 skipped एवं tokens, one skipped एव, seven repeated वगैरह tokens, and one repeated वगै0. After the correction, एवं has 14 direct-written occurrences and no suffix-adjusted occurrences. No full-run record exhausts its eligible tokens during the scan.

This remains an explicit heuristic, not adjudicated truth. The result should be evaluated as a provisional aggregate name-pattern vocabulary.

Relation to the ration vocabulary¶

The land and ration aggregates share 46,417 exact normalized token forms. Of the 28,854 ration forms observed at least 10 times, 15,858 occur exactly in the land vocabulary and cover 77,969,914 ration-member token selections. That is 98.86% of selections represented by the frequency-filtered ration forms.

Exact cross-source presence is useful anchor-vocabulary coverage, not paired evidence that a neighboring spelling should map to that anchor. The existing edit-neighbor table contains 23,612 pairs where both spellings occur in land records, 17,249 where exactly one does, and 13,320 where neither does. Upnaam does not turn those counts into canonical decisions.