The "never 100%" rule in the multi-AI arena

The Multi-AI Arena has one rule that looks like a curiosity and is actually the whole point: accuracy_pct is hard-capped at 99. In three places.

1. The prompt. The judge is told, in plain language, that 100% confidence is not available to it.
2. The parser. If the judge's JSON output says 100, the parser silently clamps it to 99.
3. The schema. The Pydantic model rejects values above 99 at the validation layer.

Three layers of guardrails for one constant. That feels like overkill until you watch what happens without it.

What happens without the cap

Models — every one of them — will, given the chance, anchor on 100%. Especially when:

• One side cites three URLs and the other cites one.
• The question is "factual" rather than "interpretive."
• The prior arena history strongly favors one side.

They don't mean certainty. They mean "high confidence relative to my available evidence." The number 100, though, has a different connotation in human reading. It says: you can stop checking.

You cannot ever stop checking. That's the actual rule. The cap is the codification of it.

What the cap costs

A model that has to reserve one percentage point of doubt has to, somewhere in its reasoning, articulate what that doubt is. That's the prize. The discrepancy analysis — what could make this wrong? — is the most useful artifact the arena produces. Without the cap, models skip past it.

What this has to do with NEWƎИ HQ

Everything I build for clients eventually gets shaken out in the arena. Pricing calls. Architecture picks. Stack choices. Outreach copy. The arena is the adjudicator when something matters and one model isn't enough.

The "never 100%" rule is not about being humble. It's about preserving the part of the answer that's actually load-bearing — the why-it-might-be-wrong section. That section is where the next move comes from.

If you want the arena installed on your own decisions, email me.