Editorial methodology

Modelwire is a curated index of AI news with original summaries and, on select stories, a short analysis layer that connects coverage to what we have already published. This page explains how that works.

Sources

We ingest from a fixed catalog of RSS feeds, official APIs (e.g. Hacker News, Reddit, arXiv, GitHub), and occasional paid news APIs. Each source has an editorial weight that nudges buzz scoring; it does not guarantee placement. See Our sources for the full list.

Buzz score

Every candidate item is triaged by an AI model (Claude Haiku) with our rubric: timeliness, breadth of interest, originality, and source credibility. The model outputs a 0–100 score, which we then adjust slightly using the source weight so trusted outlets are not systematically under-ranked. Items below our publish threshold stay in our database for audit but are not shown to readers.

Summaries

Summaries are written in our own words from the headline, snippet, and metadata we have at ingest time. We do not store or republish full article bodies. The “Read full story” link on every page points to the publisher.

Editorial analysis

On higher-signal stories (typically buzz ≥ 55 with enough entities to anchor context), we run a second pass with a more capable model (Claude Sonnet). It receives a short list of similar recent Modelwire pieces ranked by embedding similarity so the analysis can reference real prior coverage instead of inventing connections. Output is checked by automated gates: length, banned hype phrases, repetition against the summary, and consistency with any cited related story. If the gates fail, we publish the lean summary only.

AI disclosure

Modelwire uses Anthropic Claude and Amazon Titan embeddings as tools. They do not replace human judgment on policy: we set thresholds, prompts, and quality rules. Where analysis appears, the article page states that it was generated by Modelwire’s editorial layer and links here.

Corrections

We correct factual errors in our own text (summaries and analysis) promptly. See Corrections policy.