Methodology

How we collect, verify, and present political policy information.

Source identification

We identify official policy documents from each party's website, parliament.nz, and verified press releases. We do not use media reporting as a primary source.

Content extraction

Policy pages are scraped respecting robots.txt. Relevant text is extracted, navigation and boilerplate are removed, and content is chunked for processing.

AI summarisation

Each chunk is processed by GPT-4 using a strict prompt that requires neutral language, factual extraction only, and no speculation. The model assigns a confidence score.

Human review

All AI output is reviewed by a human editor before publishing. Editors verify accuracy against the source, check neutrality, and can edit or reject entries.

Publication

Approved policies are published with their source URL, source title, date, and confidence score visible to users.

Ongoing monitoring

Sources are re-checked periodically. When parties update policies, we create a new version and record the change in the Policy Tracker.

Neutrality standards

✓We describe what parties propose, not what we think of those proposals.

✓We use the same level of detail and same style for all parties.

✓We never use evaluative language (e.g. 'ambitious', 'reckless', 'smart').

✓We never infer motivation or intent beyond what is stated.

✓When a party is silent on a topic, we mark it as 'no data' rather than inferring a position.

✓We apply the same sourcing standards to all parties.

Confidence scores

Each policy entry includes a confidence score (0–100%) indicating how clearly the policy is stated in the original source.

90–100%HighPolicy is explicitly stated with clear commitments.

70–89%MediumPolicy direction is clear but some details inferred from context.

Below 70%LowPolicy is vague or inferred from general statements. Treat with caution.