Alignment at n=1: a single-subject longitudinal study of human–machine agreement, and a five-aspect model of the state it takes place in

Alignment is normally studied across populations: preferences are aggregated, benchmarks are averaged, and the unit of analysis is a distribution. This paper reports the opposite design. One person and one machine worked together over 23 days on a shared software project, with the human's rulings quoted verbatim and the machine's derived claims marked as such where the appending session marked them, in an append-only record totalling 1,272 dated entries across 2.69 MB. The machine had no memory across sessions. Continuity existed only as files, so alignment had to be re-established from zero at every boot — which made its cost and its failure modes observable each time rather than inferable in aggregate. I report three things. FIRST, A VOCABULARY IN TWO HALVES. ETHER names the state an exchange takes place in, in five aspects: environment, the world both parties are inside and neither owns; task, the request as transmitted rather than as intended; human, under-specified by default and read through the environment; expectations, the unread agreements that make the other three legible and that drift with no detector; and reflect, the operation the other four undergo. HAL names the exchange itself — human, AI, and the layer between them. I characterise that layer as a best-guess machine rather than a truth oracle, from which it follows that its correctness property is the honest typing of its own outputs and not their accuracy. SECOND, SIX FINDINGS from the record, each marked according to whether it was ruled by the subject, measured by instrumentation, or derived by the machine and left unratified. THIRD, HALT, a protocol for the case where the same problem is posed at population scale. Its relation to the first two is derivation rather than analogy: if a layer can be aligned to one person but cannot carry two — the study's central constraint — then the population-scale version cannot be a machine at all. It has to be a procedure that people run, with multi-party agreement as its gate and a human deciding at the end. HAL and ETHER describe the exchange at n=1; HALT is what is left of them when n is everyone. I state at length what this design cannot show. The load-bearing claim is the narrowest one: that a layer of this kind cannot carry two people's expectations without averaging them into a set neither party holds. A working system that does would end the argument, and section 6.1 states the mechanism precisely enough to be attacked on. --- GENERATIVE AI DISCLOSURE This paper requires an unusually specific disclosure, because the system being disclosed appears in three distinct roles and collapsing them would misrepresent the work. As the object of study: what this research studies is my working relationship with a commercially available large language model accessed through an agent harness. The model is what the paper is about. As a participant in the record: the corpus analysed here was largely written by that model under my direction, entry by entry, over the study period. Limitation 2 of the paper states this and reports the attribution split rather than summarising it away. As the drafting tool for this manuscript: the prose of this paper was generated by the same class of system, working from the record and from my rulings, and revised under my direction. Its structure, its claims, and every decision about what to include, what to withhold and how strongly to state it are mine. Accordingly: no generative system is listed as an author, and none could be — authorship entails responsibility, and responsibility cannot be held by a tool. I take full responsibility for the entire contents of this paper, irrespective of how any part of it was produced. The research itself — the record, the rulings, the findings and their limits — is human work; the machine is the instrument and the subject, not the source. --- DATA AVAILABILITY Nothing accompanies this paper. No data, no code, no appendix. A reader holds the argument and my reporting of the record, and nothing else. Section 11 states in full what is withheld, why, and what withholding costs the reader — including that the corpus is withheld in part because a living third party is named throughout it and consent for that is not mine to give on her behalf. ETHICS This is a single-subject study in which I am the sole subject, conducted on my own time, with no funder, no institution, and no external oversight body. It was not reviewed by an institutional review board because I have no affiliation through which to seek one. Section 12.2 gives the full ethics account, including the consent decision made by the one third party who is a party to what this paper reports. COMPETING INTERESTS I have no financial or professional relationship with any organisation whose products or policies are discussed here, and no institutional affiliation. One interest is declarable: the software artefact through which this method was developed is a game I am building on my own account and intend to sell, so I have a personal commercial interest in the software, though not in any finding, tool, or model named in this paper. The language model studied here is a third-party commercial product I pay to use as an ordinary customer.