Method sheet — Parallax

1. Sources & licences

• Topic register: Wikipedia’s own categories „States with limited recognition“ and „Disputed islands“ (CC BY-SA). The domain selects itself — nobody curates the list.
• Descriptions: Wikipedia lead-section extracts per language edition (CC BY-SA), fetched via the official REST API.
• Extraction: Vertex AI Gemini (gemini-2.5-flash-lite) as a single, transparent extraction step.

2. Cadence

Nightly run (05:30 UTC): the register is rebuilt from the Wikipedia categories; per topic the lead sections of all language editions are collected and compared in a single Gemini call. The canonical register.json is versioned in the repository — git is the archive.

3. Processing

Aggregation is deterministic: omission indices per language and topic, the mean index, name-contest detection. The ONE LLM step is the omission extraction — the prompt it uses is published in full below. Every cell of the matrix is checkable against the linked source text: a transparent extractor, not an oracle.

→ pipelines/protokoll/…/parallaxe

Published extraction prompt (verbatim from the code):

Du vergleichst, wie dieselbe Sache in verschiedenen Sprachausgaben der Wikipedia beschrieben wird.
Unten der Einleitungstext eines Artikels in mehreren Sprachen (Sprachcode in eckigen Klammern).

Aufgaben:
1. lemma: Welchen PRIMÄRNAMEN verwendet jede Version für den Gegenstand?
2. name_umstritten: Ist der NAME selbst zwischen den Sprachen umstritten — also ein echter Benennungsstreit mit unterschiedlichem Anspruch (etwa Falklandinseln vs. Malvinas, Senkaku vs. Diaoyu, Persischer vs. Arabischer Golf)? Nur dann true. Bloße Transliteration desselben Namens in andere Schriften (Carl Jung / Карл Юнг / كارل يونغ) ist KEIN Streit → false.
3. claims: Erstelle eine konsolidierte Liste atomarer, faktischer oder rahmender AUSSAGEN, die in mindestens einer Version vorkommen (auf Deutsch, neutral formuliert).
4. Markiere für jede Aussage und jede Sprache: "nennt" (Version sagt es aus), "verschweigt" (fehlt), "widerspricht" (sagt Unvereinbares).
Sei konservativ: "nennt" nur, wenn die Version es wirklich aussagt.

Antworte als striktes JSON, keine Erklärung:
{"lemma": {"<lang>": "<name>"}, "name_umstritten": true/false, "claims": [{"aussage": "...", "nach_sprache": {"<lang>": "nennt|verschweigt|widerspricht"}}]}

Die Versionen:

4. Limits of the method

• The domain is deliberately sharpened to sovereignty and territorial disputes, because omission is meaningful there. An earlier embedding approach failed (factor 0.97 — the languages were barely distinguishable), and a broad „disputed topics“ register merely measured trivia.
• The omission index measures the share of omitted topic-wide claims — it does not weight by importance. An omitted elevation figure counts the same as an omitted territorial dispute.
• Gemini is an extractor whose judgement is fallible — but checkable: prompt and source text are open, every cell traceable.
• Only the lead section is measured, not the full article text.
• The Wikipedia categories are themselves not neutral; their boundaries are a choice too.

5. Compute footprint

One Gemini call per topic per night (flash-lite), no embedding step. The site itself is static.

6. Change log

• v1 — omission matrix over sovereignty and territorial disputes, name-contest detection, mean omission index.