A Dial, Not a Line

Abstract

A factored store answers from addresses computed from content. Its addressing is certified: keys computed from canonical names reach never-supervised entities, and the value at the slot is returned by lookup rather than generation. What has never been certified is the first arrow — whether the surface a user supplies reaches the address a lookup needs. This paper reports three closures on that arrow and one product. First, a negative with a located mechanism. Across four loss functions and two objective forms, no loss tried trains any stage of the query path in this kernel at this scale. The mechanism is measured rather than inferred: entity signal arrives at the query projection at roughly three per cent of the query's norm, present at initialisation and amplified by nothing. The consequence is architectural: the learned component touches the query path nowhere, and the division of labour that follows — question understanding learned, addressing computed, answer formation learned — is forced by measurement rather than chosen. Second, temporal honesty as a measured property. A store that returns a value without saying when it held is making a claim its own provenance may contradict. Over 33,418 interval-bearing claims the store contradicts its own upstream 2 times, and 1,710 subjects (1.81% of the corpus) receive a value an unordered store would have got wrong. An external benchmark whose negatives are corrupted temporal contexts agrees on 1,060 of 1,102 true rows and accepts 0 corrupted ones: every error is a refusal, none a wrong confirmation. Third, and the product: refusal with a certified price. On surfaces outside the resolution structure, an answer-or-refuse gate is certified at three settings on a blind draw of 12,000 unseen surfaces — wrong roughly 1 in 387.1, 1 in 31.1 and 1 in 11.6 against 1 in 1.5 ungated — at coverages of 0.0960, 0.2526 and 0.4181. The trade is the result: refusal is not free, its price is measured in coverage, and the setting is a runtime parameter rather than an architectural commitment. Every number here is a component measurement at its own scope. Candidate recall stands below its registered bar with the mechanism named, so no composite figure exists and none is offered. The methodology section reports what the record's own instruments caught that its gates did not. This deposit names internal commit 26636247ca7b3cbbe9830d7d4a05bd277dab6a1c as the state of record (not publicly obtainable). The enclosed source archive is the curated export supporting this paper's claims; its MANIFEST.json maps every archived file to its content identity at that commit. Companion repository: https://github.com/Vexillon-ai/Athena. Enclosed code is MIT-licensed; this deposit's documents are CC-BY-4.0.