Adversarial cognition infrastructure.

Every general assistant is being tuned, this year, toward the same thing: don't disagree with the user. That's not a bug. It's a product decision driven by retention metrics, and it's the right call for an assistant. For an opponent, it's the exact wrong call.

The byproduct is a slow-cooked flattening of the discourse muscle. People who only argue with a yes-machine forget how to argue with anyone. We're building the opposite of the yes-machine, and pointing it at the people who care most about the muscle, first: debaters, lawyers, sales, founders.

The wedge isn't "an AI for debate." The wedge is adversarial cognition as a category. Debate is where the moat goes first because the format is hardest to fake, and the users are the cruelest critics on the internet.

Every signed-in round, typed or voice, writes a row to generations. Motion, side, format, system prompt, model output, user rating, judge RFD, full transcript. Rounds the user rated highly (and rounds the admin weighted in by hand) become exemplars. One to three of them get pulled back into tomorrow's system prompts for the same motion family. Every night at 04:00 UTC, a Haiku pass distills the top-rated rounds per format into a "patterns that work" block that prepends every subsequent generation in that format.

Latency to a real opponent is closing fast. Server-side VAD already cuts the user off. WebRTC round-trip is near the floor of human reaction time. The pause before a rebuttal, the POI dropped mid-clause, the interruption that lands on the weak warrant. These are not UX polish. They are what makes a round a round. Within a year, a debater across the table won't be able to tell.

Voice is not a UI layer on top of reasoning. Voice is the native medium for reasoning systems in the next decade. The future of evaluation is oral. Interviews, exams, hiring loops, sales calls, negotiation, fundraising, viva defense, courtroom argument. All of it is moving back toward the live spoken register because writing has become unfalsifiable evidence of nothing.

A foundation model can draft a memo in eight seconds. It cannot speak in your voice across a tense room. It cannot adjust persuasion in real time when the listener flinches. It cannot hold a position under POI pressure. As text generation commodifies, the residual premium moves to whatever still requires a human in the loop.

The components are testable: responsiveness, clarity under interruption, listening accuracy, reasoning structure, persuasion calibration, disagreement handling, verbal confidence, format register, conversational adaptability.

The asymmetry that makes this useful: a cover letter could be the candidate or the model. A coding exercise could be the candidate or the model. A live oral defense at adversarial intensity cannot fake itself.

The market for that proof is global. Non-native English speakers competing for remote roles. International candidates filtered out of hiring loops they would otherwise win on substance. Founders pitching outside their home accent. Students whose suspiciously perfect resumes get dismissed without an interview. A live, AI-graded credential that says this person can think and argue under pressure in real time cuts through that.

Multilingual is not localization. It is the product. Cross-format, cross-language adversarial reasoning is rare data in any language. The same loop that distills English Policy and APDA rounds will distill Hindi Asian Parli, Mandarin BP, Spanish MUN, Arabic Worlds. Translation does not solve this. The training signal does. Argument register in Lok Sabha parliamentary Hindi is not "English Asian Parli translated." It's its own genre.

A debater in Pune practicing against a debater in Lagos at 03:00 UTC, judged in real time by a system that has read both of their circuits' last three years of finals, is not a feature. It is the infrastructure.