Methodology & Sources

What this archive is, how each part was built, and what was deliberately left out — so that the work survives its makers.

What the corpus is

The Stoddard County Genealogy Archive is a free public research archive of Stoddard County, Missouri, compiled from public records. Its spine is the Stoddard County Genealogy Corpus, measured 2026-10-04 (data edition 2026-10-04):

• 228,650 newspaper life-event rows — 220,536 from Stoddard County papers (Dexter Statesman 1911–1923 capture complete: 52 weekly issues, 421 canonical pages, January 5 to December 28, 1923; Bloomfield Vindicator 1887 batch of 2,442 rows, plus backfill batches), with 4,553 neighboring-area and 3,561 regional rows alongside.
• 258,136-entry unified name index — normalized names, birth/death dates, first/last event dates, variant spellings, record counts, sources, and townships or places seen, built over the whole corpus (16 columns; v3 validated with a 9-check battery on 2026-09-28).
• 47,557-entry OCR name index (V2) — per-batch OCR name extraction, re-cleaned, deduplicated, and re-sorted after every merge, with entry-to-paper/date/page provenance for 22,082 lines and an append-only audit log of cleaner drops.
• 17,664 Missouri death certificates (1910–1975) — hand-transcribed; lane closed by Aj on 2026-09-27. (Missouri county death records cover only 1883–86 before 1910 — earlier deaths are absent by construction, not by omission.)
• ∼204,000 census transcript lines (1850–1950) — hand-transcribed decade files from 1850 (4,227 lines) through 1950 (33,725 lines). The 1890 federal census burned; only the 1890 veterans schedule survives.
• ∼63,000 burial memorial rows across 126 cemeteries — parsed Find a Grave memorials, cross-checked against BillionGraves pulls where one exists, with a conflicts file, an arbitration queue, and cross-cemetery duplicate tracking.
• 114 court-judgment transcripts (1923–2000 criminal; 1859–1990 civil) — transcribed from Caselaw Access Project reporter scans; 6 opinions not freely available are logged as missing rather than skipped silently.
• BIOENGINE entity-resolution outputs — a county run over 27,987 surnames producing per-surname candidate links with confidence bands and evidence lines; every sub-0.75 link is held for Aj’s arbitration.

How each part was built

Newspapers

Page images were captured from newspaper archives (the Dexter Statesman 1911–1923 run is complete; 1924 and later is Aj’s own work by hand). Vision readers — human-directed, plus a second extraction lane — pulled life events from each page: date, paper, clip ID, publication citation, primary and other persons, event type, place, detail, quote snippet, and confidence (11 columns). Every merge passed through merge_done_batches.py and the merge guard merge_guard.py, and a timestamped backup was written before every append (backup-first). Clip IDs such as b31-18870618-p3 are never reused across lanes.

Name indexes

The OCR name index (V2) came from per-batch OCR name-extraction runs, merged by merge_batch_v2.py and cleaned by clean_v2.py, which re-runs over the whole file after every merge — a self-healing, rule-faithful cleaner (verified 2026-10-01). The unified name index (v3, 258,136 entries) was built over the full corpus and carries Aj’s arbitration decisions in its variant rows (for example, solen → solon).

Death certificates

Hand-transcribed from certificate images into a 17-column spreadsheet: cert number, name as displayed, dates, ages, birthplace, parents’, spouse’s and informant’s names, burial place, cause of death, and uncertainty notes. The lane was closed by Aj on 2026-09-27.

Census transcripts

Hand-transcribed from the manuscript schedules for every decade 1850 through 1950, byte-verified mirrors kept on Aj’s shared drive (standing rule: intake folders live inside the drive’s census/ folder). The 1890 veterans schedule stands in for the burned 1890 census.

Cemeteries

Find a Grave cemetery intake waves were pulled through a browser (free access only) and parsed into per-cemetery folders. Where a BillionGraves pull existed, memorials were cross-checked against it. Same-person/same-date duplicates were weeded with their reasons kept; same-name/different-date pairs were never silently weeded — they went to the conflicts file for Aj’s ruling. Proposed family links (41,203, with confidence band and provenance) and 673 relationship clusters sit alongside.

Court judgments

Stoddard County criminal and civil judgments were transcribed from Caselaw Access Project reporter scans in a fixed format, with origin tagging (Stoddard-origin vs. mention-only) and per-folder indexes. 892 judgments mention “Stoddard County” on CaseMine, but anonymous users see only the top 20 per query — the rest sit behind the login wall, uncollected; the six uncollectable opinions are logged as missing.

BIOENGINE

An evidence-graded entity-resolution engine (grades A, primary sources, through D, public trees). Aj signed off on the 0.75 auto-merge threshold plus a 0.15 convergence bonus for three or more independent evidence lines. Contradictions veto merges; the corpus databases are read-only for the engine; a 29-of-29 hand-verified pilot found zero false merges. A household-level second-pass linkage is in progress.

Standing rules

• Free research only. Paid transcription credits — Archion or anywhere equivalent — are never touched; free OCR and transcription only.
• Never auto-merge ambiguous identities. Open questions go to Aj’s arbitration queue, and his word on family identity is final.
• Preserve, don’t smooth. Spelling variants, contradictions between sources, malformed dates, uncertainty, and provenance are kept as found.
• Merge guard + backup-first. Every newspaper extraction merge runs through the guard (which strips process labels, collapses mechanical duplicates, and flags blank-person rows), and every append is preceded by a timestamped backup.
• No BillionGraves cross-check without a pull. When a cemetery has no BillionGraves pull, no cross-check is attempted and intake is never blocked on it.
• Newspaper images preserved. Page images are archived and never deleted unless Aj authorizes it.

What’s excluded

PUBLICFAMTREE is quarantined. Ancestry public member-tree screenshots collected for Stoddard County roots (browser scan on Aj’s own signed-in session, read-only, 2026-09-28) are marked NEVER_MERGE: nothing from them is ever merged into the newspaper event rows, census transcripts, or any database. They serve as a verification layer for a probability engine only — a confidence signal, never ground truth.

Credit

This archive was compiled by Anthony “Aj” Jackson and is offered as a free public research resource. Every ruling on ambiguous identities is his.

Methodology condensed from the corpus data dictionary, edition 2026-10-04. Counts were measured, not estimated.