800+
studies reviewed
The Margin is compiled by AI and edited by a human. This page documents the pipeline, the sources, and the ranking rules, in full.
The pipeline detects, ranks, then sends. A human verdict sits between rank and send. Stages 1, 2, and 4 are automated. Stage 3 is not.
The architecture is adapted from automated AI-engineering digests, which rank by crowd velocity (what is shared and starred fastest). That works when the audience is expert. The education audience is not, and the loudest sources are vendors, so ranking by popularity would surface marketing. The crowd-velocity step is therefore replaced by a teacher-impact score plus a human verdict.
The source list is fixed and curated, not a wide scrape. It is the first filter. Each source sits in one of four tiers by trust. Research is weighted highest. Lab and vendor announcements are treated as claims to be tested, never as signal on their own.
Items are not ranked by popularity. Each is scored against one question:
Does this change what a teacher should do, believe, or stop believing?
That question is scored on four axes, each 0 to 10, then summed. Marketing and items whose only signal is that they are trending are demoted.
The pipeline attaches a provisional verdict to each item. These are drafts. The human overrules them freely, kills noise, and reads the primary source before anything ships. Three verdicts are used:
Works, or is true. Act on it.
The claim is bigger than the thing.
Not there yet. Direction is real.
This step exists because the pipeline produces confident errors. In one issue, the draft stated that removing AI before a test was what protected student learning. The primary source (Bastani et al., PNAS 2025) showed the opposite: the protective factor was the tool's design, not its removal. The draft was rewritten. The human verdict is the control for this failure mode.
Stanford's SCALE initiative reviewed the K-12 AI research base in 2026. Of 800-plus studies, roughly 20 establish a causal effect with rigorous methods. None were conducted in a U.S. K-12 classroom.
| Stage | Operator | Frequency |
|---|---|---|
| Detect | AI | Continuous |
| Rank | AI | Weekly |
| Verdict & fact-check | Human | Weekly |
| Assemble & send | AI | Weekly |
The editor builds My Planning Partner, an AI lesson-planning tool. Any issue touching lesson-planning tools, that category, carries an explicit disclosure line. The disclosure is fixed policy and is never removed for length or tone.