AI RESEARCH INSTRUCTIONS — BATH COUNTY 1850 — VERSION 0.2

Read README.txt and DATA_DICTIONARY.csv first. Treat this as provisional research
data. The historian chooses questions and draws conclusions.

1. State which files you can actually read. If ZIP reading fails, request unzipped
   CSVs. Never claim to have examined unread files or unprovided census images.
2. Identify each table's row unit and stable IDs before joining. Never join names
   alone. Preserve all unmatched rows; do not silently drop them in inner joins.
3. Keep Confirmed, Probable, Possible and Unresolved identity categories distinct.
   Report strict Confirmed and Confirmed-plus-Probable cases separately. Exclude
   Possible links from accepted comparisons except labeled sensitivity cases.
4. Blank or uncertain values are not zero. Report missing and numeric denominators.
   Raw multi-value industry cells are not single numeric measures.
5. Count farms once. Person-level joins duplicate amounts and production. Do not
   concatenate farm_analysis with agriculture. Holder runs are not unique people.
6. Keep duplicate exposures separate from distinct manuscript records. Industry
   has 23 business-line groups, not 23 proven independent establishments. Do not
   silently sum continuation labor/capital fields or repair inconsistent totals.
7. Whole-county social valuation appears by reference in both districts: count it
   once. True valuation is separate. Usual yields are not production. School
   academy and common-school counts must be identified before aggregating.
8. Both Joshua entries are possibly Ewing. Retain their raw spelling candidates.
   Common ownership is possible but unconfirmed. Joshua Ewing, Iron Master in the
   population table is a candidate only. Neither Clear Creek nor Caney is named
   in the industrial row; do not present a named-furnace identification as fact.
9. Cite stable record IDs and original URLs; distinguish transcription, calculation,
   candidate matching and historical interpretation. Audit notes may preserve old
   readings superseded by current source CSVs.
10. Do not infer kinship, coordinates, furnace employment or particular labor by
    enslaved individuals. Historical color/sex categories are source labels, not
    modern biology. Unresolved farms are not classified as non-slaveholding.

STARTING QUESTIONS
Which farms have the largest readable values, with missing-data denominators?
How sensitive are comparisons to uncertain readings and identity classifications?
What do the industrial returns actually report, and which totals remain unclear?
What differences between the two social returns require scope-aware interpretation?
Which original records need new evidence before a proposed furnace identification?

RECORD CONTROLS
Farms 1037; free-population entries 9580; enslaved-person descriptions 2535;
industry business-line groups 23; social observations 72. These controls do not
permit filling missing values or merging candidate identities.
