Aphanotechne: Operational Opacity and the Formal Structure of Symbolic Command

Abstract

Users of machine learning systems routinely issue a symbolically formulated request to a system they cannot inspect, and accept an output whose derivation they cannot reconstruct. That position has no settled name, and metaphor — the "black box", "modern magic" — has substituted for analysis. This article gives it formal content. It first distinguishes four irreducible modes of operational opacity — unsurveyability, untraceability, unverifiability, and unattestability — each anchored to a documented case and arising from a distinct, wholly mundane source. It then proposes aphanotechne (Gr. aphanēs, invisible + technē, craft) as an analytical category for regimes controlling, through symbolic notation, a process closed to inspection while it runs. Seven formal invariants — notational encoding, syntactic rigidity, authorisation, invocation by name, iterative execution, containment, and operational opacity — are derived from three thinly stated founding conditions by a single principle: the unavailability of observation during execution displaces every control decision to before execution. They must obtain jointly, with one documented exception: current language models present a graded input layer at which syntactic rigidity fails while holding at every layer the operator does not touch — displacement rather than refutation. Classical talismanic practice and contemporary computation then face the same constraints despite incompatible ontological commitments, because these follow from the form of mediated access, not the nature of the addressee. Nothing is claimed about the efficacy of either. The concluding argument is diagnostic: where opacity makes reason-based accountability unavailable, normative weight shifts to authorisation and containment, whose ground Thompson's attestation result shows to be institutional rather than technical.