SABERISTIC
prototypeSynthetic data only

Borrowed Brain

A structured AI decision room where distinct thinking styles interrogate one problem, debate their assumptions, and assemble a decision board.

THE QUESTION

Why this exists

A generic assistant often compresses a difficult choice into one smooth answer. People making decisions need distinct reasoning systems that expose assumptions, trade-offs, disagreements, and reversible next steps.

THE BUILD

What this prototype explores

Borrowed Brain starts from a different premise than a normal chatbot: a hard decision rarely needs one more fluent answer; it needs several incompatible ways of thinking. The visitor states a problem, adds context, and seats up to five explicit reasoning styles around a virtual table. Each brain asks the question most likely to change its recommendation, forms an independent position, challenges a real disagreement, and gets a final chance to update. The orchestration is the product. A Skeptic hunts for load-bearing assumptions, an Operator favors reversible action, an Investor prices opportunity cost, a Scientist asks for falsifiable evidence, and other brains carry equally explicit rules and blind spots. Schema-constrained model calls produce interrogation items, positions, challenges, updated recommendations, and a decision board that separates agreement, disagreement, uncertainty, and the next useful test. The visitor still records the decision and can schedule a later outcome review. The public prototype is approved only as a sample-data experience. Text entered into the flow reaches the Lovable AI gateway, browser sessions are stored locally, and signed-in save or share flows can persist records. Use one of the fictional prompts provided by the interface, treat every output as a thinking aid rather than advice, and keep real sensitive context out of the application.

DECISIONS

What shaped the build

  1. Make thinking styles executable

    Each brain carries explicit priorities, beliefs, decision rules, characteristic questions, blind spots, change-of-mind conditions, risk tolerance, evidence orientation, and time horizon.

  2. Stage disagreement before synthesis

    The flow moves from context and one question per brain to independent positions, cross-examination, revised positions, and a decision board instead of collapsing immediately into one answer.

  3. Validate every model turn

    Every AI step returns schema-validated JSON and receives one structured repair attempt when its response is malformed or violates the required shape.

  4. Keep the user as decision owner

    A session begins in browser storage and can later become an authenticated saved decision with a user choice, confidence, review date, outcome, and configurable sharing boundary.

KNOWN LIMITS

What is not being claimed

  • The sample-data approval does not cover real personal, confidential, medical, legal, employment, relationship, or financial decisions.
  • Prompts and context are transmitted to an external AI gateway, and authenticated save/share features can persist decision content in Supabase.
  • Structured disagreement can clarify assumptions, but model output remains probabilistic and may be incomplete, inconsistent, or confidently wrong.
  • The roundtable is not professional advice and must not make medical, legal, financial, employment, or other consequential decisions for a visitor.