A voice training journey built on Magnus’s second book, AI Don’t Make You Smarter — The Partnership Does. You talk to an AI counterpart, it makes a case, and exactly one thing it says is wrong. Your job is to catch it.
Currently the Stage 0 prototype — one unit, three difficulty bands, scored on verification. Nothing more gets built until that gate passes on real testers.
What it does
Each scenario contains exactly one error. It is authored, never improvised by the model — so the drill is the same drill for everyone, and the score means something.
Visible, plausible, and load-bearing. The load-bearing band is designed to catch out people who comfortably beat the visible one.
A separate model session grades 0–3 against a published anchor table, citing the evidence it used. Not vibes.
Hold to talk. Speech recognition, reasoning and speech synthesis all run on the phone — the counterpart answers in character, out loud.
No transcript, no audio, no scores are transmitted. There is no server to transmit them to.
The prototype records its own evidence — score, whether the tester agreed, time to catch, whether the plant leaked — and exports a go/no-go report. If the scoring is arbitrary, the project stops.
Privacy
The full policy is on the privacy page. The short version:
Pricing
Questions
Not yet. It is a prototype being run against a small number of testers to decide whether the concept survives.
AI Don't Make You Smarter — The Partnership Does, Magnus's second book.
