Signal

What This Is

Signal is a weekly newsletter that translates AI research papers into clear, rigorous prose. Every claim is grounded in the source paper. No hype, no speculation, no "this changes everything." Just what the researchers did, what they found, and why it matters.

The entire pipeline — from finding papers to writing about them to checking the writing against the source — runs autonomously. This page explains how.

How Papers Are Selected

Each week, the pipeline pulls new papers from ArXiv across four categories: artificial intelligence, machine learning, computational linguistics, and statistical ML. It also checks Semantic Scholar for citation data and scans Hacker News for social signals.

Every paper is scored on two axes:

• Relevance — How important is this paper to the current AI conversation? Scored using keyword matching against a weekly anchor document, institutional signals, citation velocity, and social mentions.
• Narrative potential — Can we tell a compelling, concrete story about this paper? An LLM evaluates whether the paper has a clear problem, a surprising result, and enough detail to ground an 800–1200 word piece.

The top papers are selected with diversity constraints: no more than two papers from the same institution or topic area. The goal is a well-rounded edition, not a deep dive into one corner of the field.

The only human input in the entire system is a weekly anchor document — a short list of hot topics, declining topics, and keywords that steer what counts as "relevant" this week. Everything else is automated.

How Papers Are Written

Each selected paper goes through a three-stage pipeline:

1. Citation-grounded generation. The full paper is parsed (PDF to text), and an LLM generates an 800–1200 word piece following a fixed structure: Hook, The Problem, What They Did, The Results, Why It Matters. Every factual claim must cite a specific passage — a section number, table, figure, or the abstract. If a claim can't be grounded, it doesn't get included.

2. Adversarial verification. A second LLM reads the draft alongside the original paper and checks each claim against its cited passage. It classifies issues by severity: minor, major, or critical. Types include unsupported claims, overstatements, misrepresentations, and missing context. If any critical issue is found, or three or more major issues, the paper is dropped from the edition entirely — not patched, dropped.

3. Style checking. A rule-based pass enforces the style constitution: no banned words (revolutionary, groundbreaking, game-changing), required structure, citation density minimums, and word count limits. This keeps the writing consistent across editions.

What You Won't Find Here

Signal follows a locked style constitution. A few of the rules:

• The words "revolutionary," "groundbreaking," "breakthrough," and "game-changing" are banned.
• Every piece must include what the paper does not show — specific limitations, not hand-waving.
• No speculation beyond what the paper claims. "This could change X" only appears if the authors themselves make that argument, and even then it's hedged.
• Technical terms are either explained in plain language or avoided.

The goal is that after reading a piece, you could explain the paper to a colleague — accurately, without overselling it.