Principles

    These are not features. They are commitments — translations from philosophy to design that we believe must constrain any system that claims to serve, rather than replace, human judgment.

    Delay over immediacy

    Not every response should be instant. Some queries deserve a pause — a moment for the system to indicate that this question requires more than pattern-matching, that the answer is not sitting ready-made in a probability distribution.

    Implication: Systems may intentionally introduce waiting periods for consequential decisions, not as artificial friction, but as a signal that speed and gravity rarely coexist.

    Error visibility

    Systems should surface uncertainty and doubt rather than paper over them with confident-sounding prose. When a model is unsure, that uncertainty should be visible to the user — not hidden behind fluent language.

    Implication: Confidence scores, explicit caveats, and graduated certainty indicators become first-class interface elements, not footnotes.

    Human finality

    The last judgment must rest with a person. This is not about human-in-the-loop as a regulatory checkbox. It is about designing systems that cannot complete certain actions without genuine human deliberation.

    Implication: For high-stakes decisions, the system architecture must require authenticated, time-bounded human confirmation that cannot be automated away.

    Right to silence

    The system may say 'I don't know' — and mean it. Not as a failure state, but as a legitimate response. Some questions do not have answers. Some situations do not admit of recommendation.

    Implication: Interfaces must accommodate and normalize non-responses. The absence of output is itself a form of communication.

    Non-optimization

    Not all metrics deserve improvement. Some measures, when maximized, destroy the very thing they were meant to track. Optimization must be bounded by values that cannot themselves be optimized.

    Implication: System designers must articulate which outcomes they refuse to optimize for, and these refusals must be as legible as the targets.