Methodology

How Vested sources, processes, and presents information — and what it won't do.

Sources

The full list of feeds is at /sources, kept in sync with the active configuration on the server. Sources are weighted on a 0.0–1.0 scale that reflects authority, not popularity. First-party announcements (AI labs themselves) are weighted highest; research aggregators next; tech press lower; community signals lowest. The weights affect story prioritisation but not whether a source is included.

Aggregation pipeline

• Fetch. Every feed is polled on a schedule (10 minutes for news, longer for slower-moving sources). User-Agent identifies as the Grimoire bot.
• Deduplication. Articles are deduplicated within a 7-day rolling window using a word-shingle inverted index. Two stories about the same event from different outlets are clustered together; multiple coverage of the same announcement is grouped, not repeated.
• Classification. A regex-first taxonomy classifier scores each article against the topic list. When confidence is low or the topic is high-stakes (regulation, safety research, controversy), the system escalates to an AI-assisted classifier.

AI usage

AI does the production-line work: classification, summarisation, cluster headlines, narrative paragraphs, and the reference layer's initial drafts. Three rules govern every AI output:

• Grounding. Summaries must derive from source text. If a summary cites a fact, that fact must be in the source. The system rejects ungrounded outputs.
• Attribution. Every story links back to its source prominently. Source publication names are typeset as first-class metadata, not afterthought captions.
• Disclosure. "Why this matters" callouts are marked as analysis, not fact reporting. The boundary between summarised source material and editorial commentary is always visible.

What we don't do with AI

• No AI-generated photorealism. Hero imagery is either source-attributed or absent. AI-generated diagrams, charts, and abstract illustrations are allowed and labelled as such.
• No AI-fabricated quotes. Quotes appear only when they're in the source.
• No AI-generated commentary under our editorial voice. Analysis is human-curated; AI synthesis is captioned.

Reference layer

The glossary, beginner's guides, and comparisons under /learn/ are AI-drafted, then editorially reviewed before publication. Every entry carries a "last verified" date. Where the field is moving fast (model capabilities, pricing), the date is the canonical signal of freshness — older than three months means treat with caution.

Tools

Tools at /tools/ are formula-driven where the formula is well-defined (LLM cost, token count, comparison matrices) and explicitly opinionated where they're not (decision aids). Affiliate links, when present, are disclosed inline and aggregated on each tool's page footer.

Editorial integrity

• Sponsored content is labelled "Sponsored" on every surface — newsletter, front page, story pages.
• Affiliate links are disclosed inline and at the page footer.
• We will not recommend a tool we wouldn't use ourselves.
• We will not run sponsorships from companies whose products we have actively criticised in editorial.

Privacy and tracking

Analytics is self-hosted (no third-party trackers). We track page views, referrers, and basic conversion signals — never individual users. There is no cookie consent banner because there are no cookies that would require one. Email subscriptions are pseudonymised in storage; unsubscribe is one-click and immediate.

Corrections

If you spot an error — a misattribution, a stale fact, a hallucinated figure in a summary — email [email protected]. Verified corrections are applied within 24 hours, with a note on the affected page.

Operating entity: Auva Labs FZE. See privacy, terms, and licensing.