/verify-specifics — flag what to check before shipping

Generative AI is most confident exactly where it is least reliable: fabrication concentrates in specificity — named people, dates, statistics, citations, URLs, versions, quotes (references/anthropic-best-practices.md checklist #12; source references/course-ai-capabilities-limitations.md). This skill turns a finished draft into a to-do list of the precise claims a human or a different-lineage model must verify before it ships. It does not judge truth — it surfaces the surface area.

Gate 0 — SoftServe data boundary. If the draft contains SoftServe Internal+ or customer specifics, STOP: this runs in personal Servy. Generic/public prose only. Route real customer content to the isolated ss instance. The script itself is local and network-free, but the draft must not be customer IP.

How to run it

The work is a tested, deterministic script — RUN it, do not re-derive the extraction in tokens (regex over prose is exactly the fragile-repeatable job a script owns).

1. Point it at the draft. A file: python3 scripts/verify_specifics.py path/to/draft.md or pipe pasted prose: pbpaste | python3 scripts/verify_specifics.py Add --json for machine-readable output. Exit code is always 0 (advisory).
2. Read the grouped checklist. Claims come grouped by class — DATES, NUMBERS/STATS, URLS/DOMAINS, VERSIONS, QUOTED-CLAIMS, PROPER-NOUN NAMES, CITATIONS — de-duplicated, each as an unchecked box with a per-class hint.
3. Actually verify, don't rubber-stamp. For each item, check it against a PRIMARY source — not your own memory, which is the thing under suspicion. This is where the No-Self-Review reflex applies: open the URL, resolve the citation, confirm the date/figure at its source, find the exact quote wording, confirm the named person/org exists and the attribution is right. Correct the draft in place.
4. Report what you changed. Tell Gera which claims you verified, which you corrected, and any you could NOT confirm (flag those explicitly — an unverifiable specific is a cut-or-hedge candidate, never a ship-anyway).

When to fire it

• Any time you hand Gera prose with specifics in it and "ship/post/send" is next.
• Before a /pa-post, LinkedIn draft, briefing, or wiki page goes out.
• As the verification half after /storm-research / /deep-research synthesis, on the final written artifact (those verify sources during research; this catches specifics that crept into the prose layer).

The script is the front line; your judgment closes each box. A clean run (zero claims) means low fabrication surface — still skim for vague unsourced assertions the regex can't see. Engine: scripts/verify_specifics.py (tests: tests/verify-specifics/test_verify_specifics.py).