The Emplaced Machine

Endurance for a fielded ground robot is published as a drive-hour figure. The mission examined here—holding ground—is defined by not moving, so drive hours do not measure it. This paper proposes watt-hours per emplaced-day at a stated readiness level as the quantity in which an emplaced armed system should be specified, and computes it for a thirty-day emplacement from an eleven-parameter vector. Three results follow. First, "offline" conflates two independent properties—independence from third-party services, and not emitting—and only the first is architectural; a withdrawal test separates them, and a dependency audit of a representative fielded control stack counts the parties whose withdrawal breaks the mission. Second, an emplaced armed platform should be architected as an unattended ground sensor that has effectors rather than as a vehicle that has stopped. The resulting budget is 1.22 watt-hours per emplaced-day, against roughly 360 for a conventionally architected platform. Third, and the result that transfers outside engineering: 91 per cent of that budget is bought by identification and release judgement. Holding ground without discriminating costs 110 milliwatt-hours a day; looking first costs about eleven times as much, and across the sampled uncertainty the multiple never falls below about six and a half. Discrimination is the expensive part, so its absence is what a market supplies cheaply. The usable envelope of such a system is the intersection of five regions—where it can cue, where it can identify, where its surveyed geometry is still valid, where it is still authorised, and where the effector reaches—so a sensing range and an authority range are different quantities, and only the first moves when the sensor changes. Modality substitution can expand sensing capability but may not expand human-conferred authority, which requires the control to sit downstream of every perception path rather than inside one of them. Two further design results follow. In the idealised processing-gain model, for a fixed occupied bandwidth and authenticated payload, each doubling of message duration adds approximately 3.01 dB of processing gain; acquisition, coding and implementation losses make the tabulated delivery times lower bounds and the realised gain smaller. A denial test then shows that a stop-command channel is fail-dangerous under communications denial, whereas a periodically renewed permission-to-continue channel safes the platform after a bounded interval. The latter is not "self-deactivation" as Amended Protocol II defines that term: it does not irreversibly exhaust an essential component. It is an automatically functioning loss-of-authorisation safing mechanism. The complete itemised history of corrections, narrowed claims and source upgrades appears at Appendix A.7. Version 2.12.5 restates as targets and plans reported figures that had been given as accomplished, withdraws figures not found in the sources cited for them, narrows three statements about the reported employment of ground robots, and corrects the earlier-title note and seven cross-references; it changes no model input, equation or numerical result. Version 2.12.4 updated the treaty analysis at §51, and the passages that depend on it, to the report the Group of Governmental Experts adopted on 4 September 2026, CCW/GGE.1/2026/3. Version 2.12.3 corrected one displayed rounded value. The reproducibility archive of v2.12.2, deposited with this record as Reproducibility - The Emplaced Machine - v2.12.2.zip (on the author's site as the-emplaced-machine-reproducibility-v2.12.2.zip), remains controlling because none of v2.12.3, v2.12.4 and v2.12.5 changes any model input or code. Its SHA-256 is 464974348daded7c61a4b226ea009bc93aed390f4b4818ade327db0a46c56f3f. The archive contains MANIFEST.sha256, listing the digest of every file including the generated figures, verifiable with sha256sum -c MANIFEST.sha256. Engineering preprint; not peer reviewed.