METHOD — the kit

The rules of this project. Short on purpose. Changed only with a journal entry saying why. Last changed: 2026-09-20.

1. What a record is

One file per thing, in records/: a source (one document), a person (one identity hypothesis), a family (one couple and the filiation of their children), a question (a conflict or an unknown), a prior (something believed before a document said so). Sources live in records/sources/, people in records/people/, families in records/families/, questions in records/questions/, priors in records/priors/. Dated journal entries in journal/. Templates in docs/templates/.

IDs — S-0001, P-0001, F-0001, Q-0001, PK-0001 — are assigned once, in sequence, and never reused or renumbered. F-0001 is the starting marriage. A merged-away person keeps its file, marked merged_into:, so every old citation still resolves.

Filiation is written in one place only: the family file’s children:. A person file never states its parents. (E013: a child of two families; E014: a child of their own family)

YAML traps. Quote any value that contains a comma or a colon — places especially: place: "St-Cyprien, Napierville", never bare (unquoted, the part after the comma becomes a stray key; W004). Never use bare on/off/yes/no as a key or a value — YAML reads them as booleans. Dates are YYYY-MM-DD. sighted is null or { by: Who, date: YYYY-MM-DD }. Lists are [a, b] or indented - rows, never a bare scalar (mentions: P-0001 is wrong; mentions: [P-0001] is right).

2. Grades

Proven · Probable-Strong · Probable · Possible · Unproven · Disproven

Sources are classified original (first recording, in original form), derivative (a copy, extract or transcription), or authored (a compiled work); their information is primary (recorded at or near the event by someone who knew) or secondary. A source is sighted when it has been read directly — the act image, the répertoire page.

• Compiled never exceeds Probable. A répertoire, a PRDH family page, and an online index entry (type: index — LAFRANCE at Généalogie Québec, added 21 Sept 2026) are all compiled: leads to the act, never the act. A claim resting only on derivative or authored sources cannot be graded Proven or Probable-Strong. (E006)
• A Proven claim may not cite a source nobody has read. (E007)
• A chain is as strong as its weakest joint. Chain grades are computed by build.py index and never typed. (PEDIGREE.md)
• Grade level does not decide what is printable. That is a separate, later question.

The Drouin two-part test. Au greffe is not “a later copy”: from 1679 every parish kept two registers, both written by the priest, one deposited annually with the court. A greffe volume is a duplicate original. What detects a transcription is the hand: uniform copperplate, machine-ruled paper, sequential pagination, one hand across many years → grade derivative. Several priests’ hands in one opening, strike-corrections certified in the act, marginal headwords in the entry hand → contemporaneous, grade original. Record what you saw in hand:.

3. Rules

1. Identity is a claim. The same name in two documents is two person files until a document or a stated argument links them. Look for the relational qualifier — père, fils, a stated filiation, an age — never the name. Merging and splitting are logged events with reasons. To merge: the loser gets status: merged and merged_into: <survivor>, with the reason in its body; the survivor gets a same-person-as claim citing the evidence; and every source whose mentions: names the loser also adds the survivor — otherwise the survivor’s claims citing those sources fail E008. merged_into points at the survivor, never at itself or at another merged-away file. (E009, E010)
2. Mine both sides of every act. A marriage act names at least six people. Every one of them goes in mentions: or in “Names in this act without a person file”. (W001)
3. Check before calling anything new. python3 tools/build.py who "<name>" for every name and place, before looking and again before writing. Anything returned is read and either cited or explicitly ruled irrelevant.
4. Never cite a source you have not opened. (E007; E004 catches an ID that does not exist at all)
5. Count before pattern. A bare count of same-named people is never written in prose. Rosters are generated, and carry their population and date.
6. Document-in-hand beats reasoning. Check the pixel dimensions; re-crop every proper name at magnification before typing it.
7. Grade first, prose second. Write the YAML claim, then describe it.
8. The no-peeking rule. Anything from memory, the GEDCOM, family lore or another tree is a prior, unverified. It may direct a search. It may never be cited as a source.
9. Collaterals earn their file. A witness or sibling gets a person file only to serve an identity argument or a question — on_ascent: false, and say why. (E011)
10. Errors are logged in place, attributed, never deleted. Struck text stays. Journal entries are never rewritten after their day; a correction is a new entry tagged correction.
11. Rulings are Steve’s. A question changes status only on his say-so, recorded in ruled_by:.
12. More, not less. Every session ends with a journal entry that includes what was found and not used, and what broke.

4. A session

Open: python3 tools/build.py status — the frontier. Pick a target.

Close: journal committed, check passing, status printed. That is the handoff.

5. What the build enforces

Commands: check [--skip-images] · who <tokens> · status · index [--check]. check never raises: a wrong-shaped field is E002. Every command prints files that failed to load (E001) to stderr; index refuses to write while any exist. The pre-commit hook blocks on errors or failing tests and warns when generated/ is stale.

6. Images

Act images from the Fonds Drouin (via Généalogie Québec) are for internal use only and are never committed or published (Steve’s ruling, 20 Sept 2026). They live in images/, git-ignored, named by source ID, with their sha256 in the S- record. Back images/ up to the NAS. A clone without them runs check --skip-images.

7. Sources, and what they cost

Subscriptions are highly preferable: PRDH (every Catholic act to 1849/1861, indexed by family) and Généalogie Québec — the Drouin Institute’s own site, which holds the Fonds Drouin act images and leads into PRDH. (Corrected 21 Sept 2026: an earlier text said « Drouin via Ancestry »; Steve does not use Ancestry.) FamilySearch and BAnQ are free and are recorded as such. Every S- record’s repository: says which, so the journal can say what a subscription actually bought.