How it works

The scoring engine

This page covers how 4DA scores the developer news it reads. Dependency findings (Preemption, the Brief's action items, the MCP server's answers) do not go through it: they come from your lockfiles matched against OSV advisories. For news, 4DA's job is subtraction. It reads a lot and shows you a little — only the items that clear a gate built from your own code. Here's the pipeline, end to end.

The pipeline

Your codebase                     20+ sources
  (ACE scanner + Git watch)        (background adapters)
        \                             /
         v                           v
        +-------------------------------+
        |   4-axis scoring + gate       |
        |   2+ signals to survive       |
        +---------------+---------------+
                        |
                        v
        +-------------------------------+
        |   12 quality multipliers      |
        |   depth, novelty, coherence…  |
        +---------------+---------------+
                        |
                        v
        +-------------------------------+
        |   LLM verification (top items)|
        |   strict 1–5 relevance rubric |
        +---------------+---------------+
                        |
                        v
                What survived

The confirmation gate

Every item is scored on four independent axes — semantic context, your interests, real-time signals from your Git activity, and direct dependency matches. A fifth, learned axis is reserved (held out of scoring). An item must pass 2 or more to surface.

Single-axis matches are hard-capped at 28%. No matter how strong one signal is, it cannot pass alone. This is what makes the feed hard to game: keyword-stuffing hits one axis, and without matching your codebase, your installed packages, and your recent work, the gate rejects it. Your Cargo.lock doesn't lie.

Full breakdown: The scoring axes.

Quality multipliers

What clears the gate then runs through 12 multipliers — content depth, novelty detection (introductory content is demoted; new releases and security advisories are boosted), title–body coherence, competing-tech penalties, and intent scoring from your recent work. Every constant is calibrated across 9 simulated developer personas with 245 labeled test items.

LLM verification

After keyword scoring, an optional LLM layer verifies the top items against your full context — stack, dependencies, recent commits, anti-technologies, engagement history — on a strict 1–5 rubric:

5 — must-read: a security alert for your dependency; a breaking change you must act on
3 — worth knowing: a tool that fits your exact stack
1 — noise: mentions your tech but isn't actionable

This is where the gold surfaces: articles the keyword pipeline misses because there's no literal overlap, but the model understands the conceptual relevance to your project. You run this on your own compute — local via Ollama, or your own cloud key.

Anti-gaming

Content built to score still can't win. Titles must deliver on what they promise (claim "React + Rust + Tauri" but only cover React → penalty). Repeating a keyword concentrates and hurts the score. Gamed articles get dismissed; sources that produce dismissed content lose reputation. Gaming becomes self-defeating.

Next: The scoring axes or Sources.