This repository has been archived on 2026-10-19. You can view files and clone it. You cannot open issues or pull requests or push a commit.
Files
DB-cleanup/merge-output/batches/WRITER-PROMPT.md
T
StellarCrowandClaude Fable 5.1 78e8a354d0 feat(verify): batch 01 verified (355, 354, 295, 241, 217, 216, 215, 214, 213, 212)
Ten summaries rewritten from the dossiers, fact-checked by a read-only checker
(one fair flag fixed), scenario sequences and flag proposals recorded, 16
contradictions logged for Sofia in batches/contradictions-01.md. Procedure in
VERIFICATION-PROCEDURE.md; apply_batch.py applies a batch to the stores;
batches/WRITER-PROMPT.md drives the writer agents for later batches.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GZZENdTLzNGsbNy4DyF1yt
2026-09-30 10:51:44 +02:00

41 lines
3.1 KiB
Markdown

You are verifying cases of the Gart K&R crypto-attack database. Work ONLY inside
/home/disi/Documents/Claude/GART/03-research-and-publishing/DB-cleanup/merge-output
Read first, in this order:
1. VERIFICATION-PROCEDURE.md (the seven steps and the output schema)
2. DEFINITIONS.md (the eleven scenario labels, the six violence categories, the owner's detention rulings)
3. batches/batch-01.json (the exemplar: ten finished case objects; match its depth, hedging and tone)
Then, for each case id in your list:
- Open the dossier sources/<id>_*.txt (glob on the zero-padded three-digit id) and read ALL of it:
the DATABASE RECORD block, the site's AI SUMMARY (bulk text, unverified, do not trust it), every
ARTICLE and FETCHED ARTICLE block in any language, and the OTHER LINKED SOURCES list.
- Do not use a block marked OFF-CASE, and check auto-trim-review.md: a block listed there as
OFF-CASE? for this case must not be used either. Round-up articles: use only the passage about this case.
- Write the case object exactly in the schema of batch-01.json. Rules that are not negotiable:
* Every claim in "summary" must be traceable to a line in that file. No outside knowledge, no
guessing, no "likely". Keep the source's hedges (allegedly, police said, according to).
* Name a person only if a linked source names them.
* "Violence Used" opens with "Detention — <manner>" or "No detention — <why>" per DEFINITIONS.md
(forced entry then beating or holding = detention, however short; anyone held to execute the
theft counts; a plot foiled before contact or an attempt fought off in the street = no detention).
* Record-only cases (no usable article): restructure the record faithfully, set record_only true,
and say in the summary that no article is on file.
* Non-English articles: read them and write the summary in English; never paste translations.
* "sequence": the ordered list of labels the file supports (DEFINITIONS.md names). Conventions:
foiled attack = ["Dodged Attack"] plus "attempted": [what was attempted]; a victim lured to and
held at the attackers' premises = "Kidnapping" plus a gap note. Put unsupported v1 labels in
v1_label_supported: false with the reason in rationale.
* "flags": for each of the six, record value, supported true/false, evidence line; "proposed"
only when unsupported.
* "contradictions": every disagreement between record and article, with the quoted article line
and the source name. Never resolve them.
* "facts": venue_of_control (victim's private space | victim's workplace | attacker's or
third-party space | vehicle in motion | public space | none), held_duration (none | minutes |
hours | days), outcome (released | escaped | rescued | killed | foiled | unharmed), arrests
(none | some | all | convicted). Null when the file does not say.
Write the finished array to batches/batch-NN.json (NN = your batch number) and nothing else.
Do not edit any other file. Do not fetch anything from the web. When done, reply with one line:
the batch number, the ids finished, and any case you could not finish and why.