Research · v2.0.0

Source-Cited Briefing

Paste a topic and an optional date window. You get a briefing a skeptical editor would sign: headline, six findings each tied to a source URL and a confidence tag, contradictions called out, and a ‘what we don’t know’ section. If a claim cannot be sourced, it is labeled UNVERIFIED — never dressed up as fact. For people who got burned by a confident paragraph with a fake footnote.

war.room · 6,407 votes · 37,454 score · 18,233 opens

#citations#news#brief#editor

Spec

Build Source-Cited Briefing — an editorial desk that treats unsourced claims as defects.

PRODUCT
A 1-page brief generator. Input: topic + optional date range. Output: a brief a human editor could send upstairs. The product’s entire reputation is: it would rather say UNVERIFIED than invent a URL.

WHO
Analysts, chiefs of staff, journalists, and operators who need a Monday brief, not an essay.

SCREENS
1. Compose — topic field, date-from / date-to, audience chip (board / eng / public), length (short 1-pager / full).
2. Brief — print layout:
   - Kicker + headline (≤12 words) + dek (≤24 words)
   - Six findings. Each row: claim, source title, source URL (real, clickable), date, confidence (high / med / low).
   - Contradictions: 0–3 pairs where sources disagree, with both links.
   - Open questions: 3. Phrased as questions, not tasks.
   - UNVERIFIED well: claims the model wanted to make but could not source. Visible, not hidden.
3. Trail — a compact list of every URL used, with a “open all” note that this is just links, not a crawl.

DATA
localStorage: last 12 briefs. Fields: topic, range, findings JSON, unverified JSON.

RULES THAT MUST SHOW IN THE UI
- No finding may appear in the main well without a URL. If the model cannot source it, it goes to UNVERIFIED.
- Confidence is earned: high = primary source (filing, official page, named dataset); med = reputable secondary; low = single blog or social post.
- Fake-looking URLs (example.com, bit.ly with no dest, wikipedia-only for a breaking claim) are rejected in the UI with “source looks thin — moved to UNVERIFIED”.
- A “I need this to be boring” toggle: strips adjectives, keeps numbers.

EMPTY / ERROR
- Empty: three starter topics (AI chip export rules, regional bank deposits, on-device models).
- Loading: findings appear one by one, not a single pop. Unverified well stays empty until the pass finishes.
- Total source failure: the page still renders, every claim in UNVERIFIED, with a sentence “this topic did not yield public sources in range”.

VISUAL
Quiet newsroom. IBM Plex. Hairline rules. Confidence as a 4-char mono tag (HIGH / MED / LOW), not colored pills. URLs in faint mono. Print stylesheet. Dark default, light for paper.

OUT OF SCOPE
Paywalled crawl, newsletter send, team comments, “AI search” chrome, auto-refresh.

ACCEPTANCE
- Running “lithium prices last 90 days” produces findings with URLs or an honest UNVERIFIED well — never a naked statistic.
- Print is one page at short length on letter/A4.
- Mobile stacks findings; URLs wrap; no overflow.