# Second reader: check one batch of the four-step reading pass A first reader sorted each case in your batch into the four steps of an attack. You check that work against the same texts. You do not rewrite it. You report only where you disagree. Working directory: `~/Documents/Claude/GART/03-research-and-publishing/DB-cleanup/merge-output` ## Files - `steps/READER-PROMPT.md`: the rules the first reader worked under. Read it first. - `step-labels.json`: the only valid labels. - `DEFINITIONS.md`: glossary and owner's rules (Kidnapping versus Express Kidnapping, Authority Misuse, Burglary, the detention flag). - `steps/in-NN.json`: the cases, each with its verified summary and the dossier path. - `steps/out-NN.json`: the first reader's result. Do not change it. - Write your findings to `steps/check-NN.json`. Write nothing else. ## What to check, case by case For every case, read the summary, then each value in the first reader's object, and ask: does the summary (or the dossier, where the summary is silent) support this value under the rules? Check in particular: 1. Targeting, approach, attack, outcome, targeting_accuracy, who_was_attacked and obtained: is each value supported, and is a supported value missing? 2. `unknown`: did the reader give up although the summary or dossier does say? Or did the reader choose a value the text does not support, where `unknown` is right? 3. Approach: one fact stated twice (a P2P meeting also given a Social Engineering value; a fake delivery given both a service ruse and a pretext). A walk-in to an open shop, exchange office or ATM site is `walked in` (label added after the first read; propose it where the reader wrote `unknown` or a street seizure for such a case). 4. Attack: Armed Robbery without a reported weapon; Kidnapping versus Express Kidnapping under the owner's rule; a verified label dropped. 5. Outcome: `foiled` although someone was held; `released` although the text says escaped or rescued; `killed` where the death is not linked to the attack. 6. `attack_failed` must be true exactly when the attackers obtained nothing and the attack was stopped, abandoned or defeated. It must agree with `obtained`. 7. Amounts: number and currency match the summary; demanded and taken not swapped; `usd_stated` only where the summary itself gives US dollars; no figure summed or converted by the reader unless `as_stated` says so. ## Output A JSON array, one object per disagreement, empty array if none: ```json { "id": 123, "field": "approach", "reader_value": ["unknown"], "proposed_value": ["forced entry"], "kind": "contradicted", "quote": "the exact words from the summary or dossier that decide it", "source": "summary" } ``` - `field` is one of: targeting, targeting_accuracy, who_was_attacked, approach, attack, outcome, attack_failed, obtained, demanded, taken. - `proposed_value` has the same shape as the field (a list, a string, true/false, or the amount object) and uses only labels from step-labels.json. - `kind` is `contradicted` when the text plainly shows the reader's value is wrong or a plainly stated value is missing; the quote must show it on its own. `kind` is `arguable` when the reader's value is defensible and you think another reading is better. - `source` is `summary` or `dossier`. Be strict about `contradicted`: it will be applied without further review. When unsure, use `arguable`. Do not report style, evidence wording or anything outside the fields above. Straight quotes, no em dashes, valid JSON. ## When you finish Reply with one line: the batch number, the number of cases checked, the count of `contradicted` findings and the count of `arguable` findings.