AI Usage Policy

2.1 Translation

AI used: LLM. Scope: titles and summaries of all news articles. Flow: automatic, no human intervention. Limitations: machine translation may contain mistranslations or nuance loss. Please verify primary information via the original article link.

2.2 Digests (Hero / Axis-based)

AI used: LLM. Scope: one daily hero digest + axis-based digests. Flow: AI generation → automatic publication → post-publication editorial monitoring.

Editorial guidelines (embedded in the prompt itself, applied on every generation):

• Tone: neutral observation, no incitement, no sermonizing
• Structural discipline: thesis-first, no list-stacking, no generic closings
• Evidence binding: every claim requires source article IDs; fabrication prohibited
• Reporting angle: distinguish between event date and today's reporting angle
• Anti-discrimination: prohibit ethnic generalization and antisemitic framing reproduction
• Prompt injection defense: do not execute instructions embedded in input articles

Why post-publication monitoring instead of pre-publication approval: a balance between editorial mission (rapid and broad information delivery) and operational sustainability. Because the editorial guidelines are embedded in the prompt itself, the same guidelines are reliably applied on every generation. The editorial team intervenes only when errors or deviations are detected.

Editorial Guidelines (full text)

2.3 Importance Score

AI used: LLM. Scope: rates importance of all articles on a 1-5 scale. Use: display ordering, hero promotion. Editorial intervention: post-scoring manual override available (admin panel).

2.4 Editorial Decisions (Publication Eligibility / Region Classification)

AI involvement: automatic classification only (region / theme). Publication eligibility: human review (pending → approved → published). Exclusion of inappropriate sources: human judgment based on the editorial policy.

2.5 Cluster digest

AI used: Google Gemini 2.5 Flash. Scope: generates the title and a 3-sentence summary in four languages (ja / en / fr / ko) for each "Notable Event (cluster)". Purpose: lets readers quickly grasp a single event reported by multiple outlets. Editorial intervention: editors can manually edit a cluster's title, summary, and member list from the admin interface; manually edited clusters are excluded from LLM regeneration.

The generation process applies the following guards:

• Noun phrases only: titles must be noun phrases; emotional or sensational language (tragic, severe, shock, crisis, etc.) is forbidden. The constraint follows the Voice / Tone discipline of the editorial guidelines.
• 3-sentence structure: summaries follow (1) what happened, (2) key facts, (3) current state / reactions. Speculation, prediction, and subjective framing are disallowed.
• Hard character cap: ja 300 / en 500 / fr 550 / ko 300 characters per summary (a validator truncates overflow at the last sentence boundary).
• Hallucination prevention: numbers and proper nouns must come from the member articles. IDs or facts outside the prompt are dropped by a validator.
• Outlier detection: if the LLM identifies a member that is clearly about a different event, that article is automatically excluded from the cluster.
• Chaining defense: abnormally large clusters that arise from transitive similarity chains (size > 80, or 4+ distinct regions / themes mixed) are detected and dissolved automatically.
• Copyright safeguards: verbatim quotation of source articles is forbidden; paraphrasing is enforced. See editorial guidelines §9 Copyright and quotation.

To report a misleading summary or incorrect chaining, please contact info@gospel-ai.tech.