Sources & Methodology

Most news sites treat their production process as a trade secret. We think it is a trust signal. Here is exactly how an article goes from a rumour on a forum to a published piece on RetroShell.

Step one: discovery

Our pipeline monitors more than 70 retro-gaming sources continuously. Sources include: specialist sites (Time Extension, Retro Dodo, Nintendo Life, Hardcore Gaming 101), general gaming outlets (Vice, Kotaku, Engadget, NotebookCheck), Google News RSS for 60 retro-specific search terms, a handpicked X/Twitter signal list of hardware modders, preservation archivists and retro journalists, and an expanding set of RSS feeds discovered automatically from the publisher domains we cite.

Step two: deduplication

The same story is often reported in parallel by five to ten outlets. Our system dedupes at four tiers: URL canonicalisation, source+title similarity (Jaccard distance at 65%), title similarity across all candidates at 65%, and TF-IDF content similarity at 80%. A Gemini batch-dedup pass catches the cases that slip through the first four tiers.

Step three: relevance

Before we spend an LLM token rewriting anything, we filter out the irrelevant: modern-gaming news, off-topic lifestyle content (astrology, crypto, dating), aggregator listicles (“Top 10 X, ranked”) and opinion essays. This is handled by a three-way content classifier that routes output into news, features or lists with different quality bars per category.

Step four: source-ladder walk

For every candidate story, we attempt to walk the link chain back to the primary source: the creator, developer, official press page, GitHub project or archive. If the feed source is an aggregator rewrite and we can identify a higher-trust primary, we attribute the story to the primary and credit the aggregator as “first reported by”. We maintain domain-specific heuristics for more than twenty major rights holders (Nintendo, Xbox, Sony, SNK, Capcom, Square Enix, Sega, Atari, Bandai Namco, Taito, Konami and others).

Step five: AI-assisted drafting

Stories that pass the filters are drafted by Google Gemini 2.5 under one of five author voice profiles (see Masthead). The draft is constrained by: no em-dashes, metric units, British spelling (except for Marcus Rivera’s NA console beat), no invented first-person experience, specific source attribution, and a minimum word count appropriate to the subject. The writer prompt includes hard rules that force the model to return newsworthiness_score: 0 and exit for speculation, opinion or feature pieces without a reportable fact.

Step six: fabrication guardian

A second Gemini pass audits the drafted article. It lists every specific personal, factual or referential claim and marks each supported, inferable or fabricated against the source material. Any article with a fabricated claim is regenerated in third-person reportage mode. If the regeneration still contains fabricated claims, the piece is dropped.

Step seven: quality gates

Every article passes through eight gates before it can publish: source-ladder trust, Devastator fact-check, fabrication guardian, voice enforcement, similarity (plagiarism) check, image validation, SEO-completeness, and link validation. Hard gates must all pass for an article to auto-publish. Any gate failure holds the article as a draft for editorial review by the Editor-in-Chief.

Step eight: publication

Articles that clear all hard gates with a newsworthiness score of 9 or higher can be published automatically up to a daily cap. Everything else waits for the Editor-in-Chief’s review. Features and lists never auto-publish; they always wait for a human editor.

AI tooling we use

  • Google Gemini 2.5 Flash and Pro: article drafting, dedup, source-ladder walking, fabrication audit.
  • xAI Grok 4 Fast: fact-check cross-reference and X/Twitter signal mining.
  • Google Nano Banana 2: generative feature images where source imagery is unavailable and clearly labelled in the image caption.
  • Anthropic Claude: audit and security review of our own pipeline.

Why we tell you this

Trust in online news is low for a reason. We think publishers should be honest about how they work, especially when AI is part of that process. If you have suggestions for how we could improve or flag a concern about the pipeline, email editor@retroshell.com.