The Honest Cutting Edge: A Keyword Floor in Front of Two Frontier Models

Everyone wants their agent to “know when it’s unsure.” Most ship a vibe — a model asked to grade its own confidence, which is exactly the judgment you couldn’t trust in the first place.

Here’s the honest version we run instead. Two tiers, no mystery:

A deterministic floor. A transparent keyword-and-structure classifier, zero model budget, that decides whether a decision needs review. Publish, deploy, delete, account/secret, global policy, any “mark verified” claim — it flags them and cannot be talked out of it. It runs even when every model lane is down.

A neural ceiling. Only when the floor flags risk do we consult two real frontier models (Gemini 3.5 + ChatGPT) over the browser lanes we already pay for. Their answers are gated deterministically: two independent approvals proceed, one explicit block escalates to a council, and a lane that’s blocked or empty is recorded as exactly that — never as agreement. The system cannot fabricate its own approval.

That’s the whole trick. Not a black box that claims to know its own limits — a plain, auditable gate that decides when to route to the strongest reviewer you can reach, and proves the answer instead of trusting it.

The unit of reliability is the gate, not the model.

Support proof-first autonomy work (sponsorship, not equity): paypal.me/exzilcalanza