# Verification procedure (per case) — the seven steps that turn a case green Written 2026-09-30. Applies to every case whose summary is not yet in summary-overrides.json (319 as of this date: 185 template-short, 134 long but never read against sources). The 70 already done are the exemplars: read two or three of them first (e.g. 248, 334, 370). Definitions: DEFINITIONS.md (the paper's glossary plus the owner's detention rulings). Hard rules from README-START-HERE.md still apply: never edit the DB export, never fabricate, never translate into the archive, log every record-vs-article contradiction instead of resolving it. ## Inputs per case - sources/_*.txt: the dossier. DATABASE RECORD block, the site's AI SUMMARY, the sheet ARTICLE blocks, the FETCHED ARTICLE blocks (KEEP = trimmed, UNTRIMMED = raw page), the OTHER LINKED SOURCES list with verdicts. - Round-up articles: use only the passage about THIS case. - OFF-CASE blocks and sources on auto-trim-review.md as OFF-CASE?: do not use their content. ## The seven steps 1. Read the record block and every article block in full, any language. 2. Write the summary in the six categories, from the file only: Victims / Attackers / Attack Method / Violence Used / Crypto Demanded/Stolen / Status. Violence Used opens with "Detention — " when anyone was held (owner's rule: forced entry followed by beating or holding, anyone held to execute the theft, any duration) or "No detention — " when the file shows the attempt never gained control. Keep the source's hedges. Name a person only if the linked source names them. Record-only cases (no usable article): restructure the record faithfully and say so. 3. Fact-check: every claim in the summary must be traceable to a line in the same file. The checker (second agent, read-only, no fetching) flags FABRICATED / CONTRADICTED / OVERSTATED. Fix every fair flag before the batch is locked. 4. Scenario sequence: the ordered list of the eleven labels the file supports, one line of rationale, plus "attempted" (for foiled attacks: what was attempted) and "gaps" (what the labels cannot express). Conventions in force: foiled = Dodged Attack + attempted[...]; lured to and held at the attackers' premises = Kidnapping with a gap note. 5. Flags: for each of the six, state whether the file supports the record's value; propose a change with the evidence line when it does not. Detention per DEFINITIONS.md. 6. Contradictions: every disagreement between record and article on date, location, victim, amount, coin, weapons, detention, violence, drugs, deaths, attackers, arrests. Quote the article line. Do not resolve. Also flag OFF-CASE links (article is a different incident). 7. Facts (optional columns): venue_of_control, held_duration, outcome, arrests, amounts in USD when the source states USD. ## Output: one JSON object per case, appended to batches/batch-NN.json { "id": 217, "summary": "**Victims:** ...\n\n**Attackers:** ...\n\n**Attack Method:** ...\n\n**Violence Used:** ...\n\n**Crypto Demanded/Stolen:** ...\n\n**Status:** ...", "sequence": ["Malicious Invitation - SocialEng", "Kidnapping"], "attempted": [], "gaps": [], "v1_label_supported": true, "rationale": "one line", "flags": {"kidnappings": {"record": 1, "supported": true, "evidence": "..."}, "...": {}}, "contradictions": [{"field": "date", "record": "March ?, 2025", "article": "...", "source": "Expressen"}], "off_case_links": [], "record_only": false, "facts": {"venue_of_control": "...", "held_duration": "...", "outcome": "...", "arrests": "..."}, "notes_for_sofia": "..." } ## Applying a batch (scripts, run from merge-output/) - apply_batch.py batches/batch-NN.json -> summary-overrides.json, scenario-sequences.json, detention-flag-corrections.json (proposals), batches/contradictions-NN.md (for the worklist) - build_cases.py --all && verify_all.py (must PASS) - export_xlsx.py (the case turns green in the export) - commit with a message naming the batch and the ids ## Batch order Descending id over the 185 template-short cases first, then the 134 long unverified ones. Ten cases per batch. Batch 1 is written by the primary session as the exemplar; later batches are produced by writer agents and checked by a read-only checker agent.