should() in 16 slides: 23 live decisions for $0.00023

This is the deck I take into rooms when someone asks what should() actually does. It follows one made-up campaign, a regional auto dealer's November buy, from the first page check to the audit trail, and every decision on it is real: 23 live calls on 12 October 2026, decision ids and all.

The animated deck. Sixteen slides, keyboard or scroll, with a play button if you'd rather sit back. Open the deck →

What's in it

• One question in, one accountable answer out. Two pages, same question, a yes at 0.88 and a no at 0.94, each with a decision id.
• Rules first, model second, your policy last. Hard limits are checked in code before any model runs. A clear breach decides alone, at $0.
• The engines. Jev, Clef, Clef-flash, GPT-6 Luna and GPT-5 mini on the same 120 deals. Jev keeps the job. The full numbers are in the case study.
• A deal gate. Offer A clears the bar. Offer B breaks the $14 CPM cap and never reaches the model.
• Ten proposals in, ten verdicts in under a second. Six settled by rules for free, one sent to a trader because the seller quoted a price range in prose.
• Where should() gates an OpenProposal. Our mapping of the gate onto IAB Tech Lab's AAMP 3.0 draft: qualify, negotiate, bind, report. Our proposal, not the spec.
• Same deal, stricter policy, different verdict. Raise the yes bar to 0.95 for anything that commits budget and the same 0.9 becomes an escalation.
• Batch, choice, score, outcomes. One weak check holds a launch. A pacing decision picks "widen supply" at a 41% win rate. A creative that dropped its disclaimer goes to review. A delivered deal gets its outcome recorded against the original decision.
• What makes it safe to put money behind. Fails closed, bars only go up, every call logged, safe to retry, measured on your own deals.

The fine print

The campaign, the brief and the sellers are invented. The decisions are not: each one was made live on 12 October 2026 and the ids on the slides are the first characters of the real ones. The engine comparison on slide 5 is synthetic, one run, engines alone, with the caveats spelled out in the case study. Timings are from a laptop.

If you'd like to run your own deals through it, request access and gate one real decision this week: a deal, a bid change or a creative.