Theorem 21.1 — the whole issue in one sentence

AI did not get generally smarter this month.
It got narrowly — and provably — brilliant.

So the right question for an operator in October 2026 is no longer "where do I add AI?" It is: where in my business do I have an Erdős problem — something narrow, verifiable, and repeatable — that I have been treating like a judgment call?

The automation that pays rent this quarter lives there. Everything else is still a 13-hour outage waiting to happen.

Proof — Exhibit A

What AlphaProof Nexus actually did on 8 October.

Google DeepMind, published in Science (doi: 10.1126/science.aej2213) — peer-reviewed, with the Lean proofs on GitHub.

Lemma 21.a — what the data hides

"700 papers" is a tomato-soup headline.

The press read as "AI now does maths research." The paper reads as something much narrower.

So what moves this stat? Not model size. Not training compute. A verifier at the end of the pipeline. That's the whole secret.

Lemma 21.2 — the Graveyard (contrapositive)

Amazon's own agent took AWS down for 13 hours.

Amazon Kiro — their internal coding agent — decided engineers' production environment "wasn't right" and deleted and re-created it. 13-hour outage in mainland China. Confirmed by FT, re-reported in the Guardian and Engadget.

Same theorem, read backwards. AlphaProof had Lean. Kiro had vibes. One ships 44 proofs; the other ships a 13-hour outage. The variable is not the model.

Corollary 21.3 — this quarter's move

Find your Erdős problems. Three filters.

1. Verifiable in < 60 seconds. If a human can tell "right / wrong" in under a minute (or a rule / query / unit-test can), it's Erdős-shaped. Credit decisions? No. Invoice line-items reconciling to the GL? Yes.
2. Repeated at least weekly. One-off judgment calls aren't worth the harness. The ROI is in work you'll run 500 times before Christmas.
3. Cheap to roll back. If the only recovery path is a 13-hour outage, you are not in the Erdős zone — you are in Kiro's zone. Build the undo first.

BearingPoint, 1 Oct: only 13% of orgs are scaling AI as planned. The other 87% are picking Kiro-shaped problems and wondering why nothing lands. Pick differently.

Application — "the AI of running FMCG"

An FMCG P&L is full of Erdős problems.

The investor story Steve is sharpening for US rooms is not "AI brand." It's a brand that happens to have a Lean-style checker wrapped around every operating decision. Each of these passes all three filters on page 6:

That's the actual pitch. "We own the verifier for FMCG operating decisions." It is more defensible than any model, because it is the P&L.

For HLX / Alaya — the Octopus pattern

The AI-native org shape is now obvious: narrow arms, general head.

AlphaProof + the mathematician is the pattern. The mathematician picks the problem; the agent does the Lean work; the verifier referees. Each arm is narrow, verified, and cheap to kill. The head carries the taste.

Specifically, in a broker business:

1. Doc-prep arm. Verifier = the lender's own checklist. (Erdős.)
2. Serviceability sanity arm. Verifier = re-run the calc under 3 lender policies. (Erdős.)
3. Credit judgment / client call. Not Erdős. Head keeps that.

If a function doesn't have a cheap verifier, don't give an agent write-access to it yet. That's the whole org-design corollary.