The guide
Practitioner instructions for doing AI-assisted quantitative work rigorously: not how to prove a theorem, how to tell when the model's answer is wrong before you ship it. Seven entries are written as a first pass below; more are still being added, and each says where it stands.
New to this · In progress
The one idea underneath everything else here: self-review catches typos, internal consistency catches arithmetic, independent re-derivation catches logic, external authority catches framing, and a hostile expert catches what you were sure of.
A model rarely hands you an answer that looks wrong. It hands you one that looks right and is not. What that failure mode actually looks like on the page.
Pre-registered anchors, kill criteria, and reproduce-before-extend: how to commit to what would count as wrong before you have a result to be attached to.
Why two independent-looking programs are not actually independent when they share a library, a convention, or an assumption, and how to hunt for the shared step.
A result checked at size n is not checked at size n+1. The specific ways floating point and off-by-one boundaries produce silently plausible wrong integers.
What belongs in an audit trail: models used, prompts, transcripts, disclosure wording, and the submission checklist that keeps a claim honest after it ships.
Turning the ladder into a shared habit: how to run a review where the goal is to break a plausible answer, not to approve it.
The public research ledger the guide is drawn from: the claims registry, the verification method attached to each entry, and the receipts.
Every entry here is drawn from an actual practice, not a curriculum written to fill a page, written up as the TSO discipline: every claim gets an anchor, every computation gets a second independent path, every mistake gets logged instead of buried. It is the same practice behind the public research ledger at Checkable, reframed here for builders instead of mathematicians. The full four-week version, with labs and a graded capstone, is the course.