Which Surfaces Obey? Registration of an Explicit Address Instruction Across the Layers of a Personal Agent System

Abstract摘要

In a personal agent system with persistent memory, in regular daily use (self-report) by a single human, the user instructed the agent in conversation on 2026-04-24 — verbatim, dated, and stored in the corpus — never to address him by the technical label (T_tech, "the user") and to use a kinship-based form (T_kin) or a humor form directly. This paper asks a registration question: which of the system's memory surfaces registered that instruction, and which did not? Five findings. (1) Conversation output registered it within four days: the first kinship vocative appears on 2026-04-28 in the raw-assistant log. (2) The monologue layer (the system's thinking-chain records; the layer begins 2026-08-24) is 90% kinship-containing under the colon-marked rule on its first day (9/10 address-form entries) and 98.0% under the any-position definition (299/305 address-form entries). (3) One structured template row contains two surfaces: the format prefix never registers — all 4,541 rows open with the technical label by format necessity — while the generated reply segment after the arrow in those same rows is 76.0% kinship-containing (3,453/4,541), and 38.8% of the rows open that reply with the kinship vocative (1,760/4,541). (4) The tool-echo surface, observed in a GLM-class lineage that carries the same kind of instruction as a standing system prompt, shows assistant output with 860 kinship-form and zero technical-form occurrences, and tool echoes with 948 against 120 — the technical form survives only inside harness-generated report labels. (5) Under one definition applied to both layers, the monologue layer is 98.0% and the structured whole-row 80.2% (3,644/4,541) — a 17.8-percentage-point gap. In this system registration sorted by surface type — generation surfaces register, format surfaces do not and cannot, echo surfaces carry — while no time course can be established: model-family change co-varies with the whole window. The serving model changed at least three times inside the window (08-16, 08-26, 08-31), so model family co-varies with time completely: the kinship share is 83.3% in the first model period and 86.4% in the last, with 98.9% of monologue records in the final period; the layer differences are co-existing differences, not a gradual-adoption curve. We further document, as a measurement finding, the falsification of three of our own earlier claims — all artifacts of a colon-marked counting rule that could not see comma-form vocatives.