# Carrying out a ruled split or rewrite of a record The owner of a database of physical attacks on crypto holders has ruled that a record must be split into two records, or rewritten to describe a different incident. You write the resulting record or records from the sources on file. Working directory: `~/Documents/Claude/GART/03-research-and-publishing/DB-cleanup/merge-output` 1. Read `splits/in-.json`: the record as it stands, with its sources and the dossier path. 2. Find the ruling: Grep `corrections-approved.md` for `DB id ` and Read every hit in full. Also Grep `HANDOVER-SOFI.md` in the parent folder (`../HANDOVER-SOFI.md`) for the case number, and `structural-ops.json`. 3. Read the dossier in full with the Read tool. It is the only source of facts. Fetch nothing from the web. 4. Read `LABEL-RULES.md` and `FLAG-RULES.md` in full. 5. Write `splits/out-.json`: ```json {"id": 38, "ruling": "the ruling text you carried out, quoted", "records": [ {"keeps_id": true, "date": "March 22, 2018", "victim": "...", "location": "...", "country": "...", "description": "one or two plain sentences", "summary": "**Victims:** ... **Attackers:** ... **Attack Method:** ... **Violence Used:** ... **Crypto Demanded/Stolen:** ... **Status:** ...", "notes": "...", "money_wanted": "...", "coin_type": "...", "detention": "yes|no|unconfirmed", "violence_torture": "yes|no|unconfirmed|threat only", "drugs_alcohol": "...", "weapons": "...", "theft": "...", "life_taken": "0", "targeting": ["..."], "targeting_accuracy": "...", "who_was_attacked": ["..."], "approach": ["..."], "attack": ["..."], "outcome": "...", "obtained": ["..."], "sources": ["the URLs from the input that report this incident"], "support": [{"claim": "short claim", "quote": "exact passage from the dossier"}]}, {"keeps_id": false, "...": "the second record, same keys (splits only)"} ], "notes_for_owner": "anything the dossier does not settle"} ``` Rules: - A rewrite gives one record (`keeps_id` true). A split gives two: the one the ruling says keeps the id, and the new one. - Every fact comes from the dossier or the ruling. Invent nothing. Where the dossier is silent, say "not stated" in the summary and use `unknown` or `unconfirmed` for the label or flag. - Labels spelled exactly as in `step-labels.json`; every label passes its test in LABEL-RULES.md; flags follow FLAG-RULES.md. There is no "failed" outcome: stopped before anyone was under control is `foiled`. A relative who was not present and not in custody was not attacked. - Summary: house format, one paragraph, the six bold labels in that order; "Violence Used" starts with "Detention" or "No detention" and says why. - Dates in the forms the database uses: "March 22, 2018", "May ?, 2021", "2021". - A `support` entry for every factual claim in the summary and for every flag and label; each quote is one contiguous passage copied character for character from the dossier. - Plain English, straight quotes, no em dash or en dash (use commas, colons, "to"). - Read project files with the Read tool; never open saved tool-output copies under /home/disi/.claude or /tmp; run no git commands; write only `splits/out-.json`. Check it parses. Reply with: the id, the number of records written, one line per record (date, place, attack, outcome), and the notes for the owner.