Compression Is Not Preservation: A Question-Relative Architecture for What a Computation May Discard

Abstract

Software discards information constantly. A log is sampled, a table is aggregated, a context window is truncated, a sensor stream is fused, a state is reduced to a latent vector, and a monitor keeps a checksum instead of the record. Some of those reductions are exact. Some can be undone. Some retain exactly what a stated question needs. Some drop a distinction that question still uses, and then a later decision treats the smaller object as if it were the original. This note proposes a test for that last step, not a new compressor. Compactness does not establish adequacy for a question Q. Exact reconstruction is a different claim from competence for Q. Competence for Q does not license every action that might be taken from the reduced state. An unresolved input is not failure of the whole computation. It restricts the answers and actions that depend on that input. The first empirical object is a comparative benchmark on declared questions, not a product claim. Pre-results methodological paper. The constructed example demonstrates the proposed failure mode. No comparative performance claim concerning an existing compression system is made. Exact competence is equality of the declared readout and the question on a declared domain. A tolerance is a separate, approximate claim. Loss relative to the source is not failure for the question. An empty dependency intersection does not license an answer, and competence for the question does not authorize an action.