Methodology

The pipeline

Every issue moves through one pipeline:

research → verification → editorial synthesis → review → merge → publication

Candidate stories are collected throughout the week, ranked (technical novelty, engineering impact, source quality, reproducibility, practical usefulness), and only the strongest few make the issue. The funnel is deliberate: many candidates, a handful of published stories, no filler.

Research material and published content are kept strictly separate. What you read on this site has passed human editorial review; the messy working material behind it never appears here.

The claim ledger

Before an issue is drafted, every checkable fact gets a row in a claim ledger recording:

  • the claim as published,
  • its source and the exact location the fact came from,
  • its evidence tier,
  • its decision cost,
  • its verification status.

If a fact is not in the ledger, it does not enter the issue. Nothing enters a draft from memory.

Evidence tiers

Tier Meaning
P-V Primary source, vendor-authored
P-I Primary source, independent
S Reputable secondary reporting
C Community claim
INF Our arithmetic inference, labelled as such
U Unverified — flagged, never asserted

Vendor-reported figures are labelled as vendor-reported every time. When independent evidence does not exist yet, we say so rather than papering over it.

Decision cost

Reliability of a source is not enough — we also grade what it costs a reader if we are wrong. HIGH decision-cost claims — licenses, prices, hardware requirements, context limits — are re-read from the primary artifact every issue, even for models we have covered before. A benchmark error costs a reader an experiment; a license error costs them a legal review.

Source tiers

Sources fall into four tiers:

  • Tier A — primary technical evidence: papers, code, model cards, technical reports, benchmark repositories, official documentation.
  • Tier B — primary reporting: engineering blogs, release announcements, researcher statements.
  • Tier C — independent verification: independent benchmarks, reproductions, external analyses.
  • Tier D — discovery: news, social posts, aggregators, newsletters.

The rule: Tier D tells us where to look. Tier A–C tell us what we can publish. Nothing is published on the strength of a discovery-tier source alone.

Pre-ship passes

Before any issue ships, separate passes check: every number re-traced to its ledger row, the week’s stories cross-read against each other, the load-bearing conclusions red-teamed against their sources, and model names and figures checked for internal consistency.

Corrections

Material factual errors are never silently fixed. Each correction ships in a dated Corrections section: what was wrong, what is correct, the primary source it was re-verified against, and the root cause. Claims we could not verify carry forward as open items to the next issue.

Humans publish

Automation may assist research; it never publishes. Every issue is approved by a person before it goes live.