Everyone can generate an answer. Not everyone can tell if it's wrong.
A practitioner lane on verifying AI-assisted quantitative work: for the PMs, engineers, and quants who use a model to get to an answer and then have to decide whether to trust it. Not a research-math course. The TSO discipline, taught by someone who runs it on the record: every claim gets an anchor, every computation gets a second independent path, every mistake gets logged instead of buried.
Guide · In progress
Self-review, internal consistency, independent re-derivation, external authority, hostile expert: the escalation ladder for catching a plausible wrong answer before it ships. The guide index has the full breakdown.
A Lightning Lesson on reproduce-before-you-extend: matching a published table to four decimal places, then the three rows that would not match, and what that turned up.
Counting by listing is exponential death. A Lightning Lesson on the bug catalogue that makes exact-arithmetic shortcuts dangerous, and the anchors that catch it.
Agreement is not proof. A write-up of the correlated-error trap: two independent programs sharing one unexamined step, and only an external anchor catching it.
Pre-registered anchors and kill criteria, written before a single line runs. The habit that turns errata into a byproduct instead of a crisis.
The public research ledger this practice runs on: the claims registry, the verification methods, and the receipts, entry by entry.
Verify Like a Mathematician is a four-week cohort on shipping AI output you can actually trust, built for technical PMs, engineers, and data people, not research mathematicians. It is in development, not yet open. The syllabus and the free preview lessons are on the course page.
Everything on this site is stated at exactly its status. In development is not open for enrollment; in progress is not published; coming soon is not written yet.