Tarski's Ladder: Deriving an Ethical Space Instead of Imposing One

When we ask what ethical considerations a population of language models appeals to, we meet a fork that is philosophical before it is technical: tag each response against a fixed catalogue of moral principles, or let the structure of the ethical space emerge from what the models actually say. This paper argues for the second on a specific ground. The structuring category must sit a level above the data it organizes, derived rather than axiomatized. This is the lesson of Tarski's undefinability theorem and Goedel's incompleteness results: a sufficiently expressive system cannot define its own truth, or its own categories, from inside; one ascends to a metalanguage. We give that lesson a second reading through object-oriented ontology: the language model is a non-human object whose reality is never exhausted by any relation to it, so its ethics cannot be read off its surface, only partially translated, a level up, from what it emits. We make both readings concrete with a small running artifact, metarung, an operator-agent rigor loop in which an informal justification climbs a ladder of formal languages (surface in TypeScript, logic in Prolog, meta in Lisp), has its ethical structure derived bottom-up, is checked for faithfulness by a cheap judge whose misses re-enter the loop, and returns as an actionable summary. We present a worked example, report what emerged, and put the open questions to philosophers directly. The contribution is not a theory of ethics. It is a method, and a demonstration that a commitment about where categories live can be carpentered into code.