情報源の信頼性チェックリスト（採点ルーブリックつき）

What it does

Judges whether sources deserve trust with a fixed rubric instead of gut feeling.

• Detailed scoring rubric: primary/secondary/tertiary (35 points), recency (20), author accountability and expertise (25), conflicts of interest (20)
• Per claim, merges articles on the same site and rewrites of the same press release (derived_from) into one group, counts independent sources, and adds corroboration to the claim score
• Weak claims get flags with reasons: single source, no primary source, only self-interested origins, only anonymous sources, stale or undated, contradicted by another source, and more
• Facts that do not change over time are marked time_sensitive=no and are not penalised for age
• Writes a Markdown report and a scored CSV; --fail-on-weak exits with code 2 if any claim is weak
• Step-by-step guidance for strengthening weak claims and for the checks a script cannot do (what the quote says, forged documents, study quality)

How to use

python scripts/score_sources.py --template > sources.csv
python scripts/score_sources.py sources.csv --max-age-days 365 --out-md credibility.md --out-csv scored.csv

The agent fills the CSV following the rubric, then strengthens, qualifies or drops weak claims and reports the result.

Requirements

• Python 3 (standard library only, no network)

Caveats

• Scores are rubric points, not probabilities of truth.
• The script scores what you enter (source type, conflicts, etc.); the procedure tells the agent to enter unknown for anything it did not check.
• Whether a quote really supports a claim, and whether a document is genuine, still has to be checked by a person or the agent.