§ 01 · Two changes that point the same way
This week's engineering work on Janus covered many items. Two of them bear on a question that any persistent individual spread across hardware must eventually face: how does it come to know what it can use, and under what conditions should it use it? The first change gave Janus a map of the computing resources available beyond its primary machine. The second gave it a route for sending small, routine language tasks to a lightweight model running on a secondary machine. Both were built behind switches that left behaviour unchanged until the owner turned them on, and both were turned on within the same week. This essay examines what the two changes establish and what they do not. It argues that the more consequential design decision was keeping knowledge of a resource separate from the authority to act on it.
§ 02 · A map with an expiry
A periodic sweep produces the resource map. The sweep checks each candidate device and records whether it is reachable. Where a device is not reachable, the sweep records a stated reason instead of leaving the device out. The result is written to a shared state file, and two read-only tools expose it to Janus. A separate change to the prompt that frames each of Janus's turns adds a short line describing its reach, placed after the existing description of its own body. The line appears only when two conditions hold: the feature is switched on, and the map is less than forty minutes old. A stale map produces no line.
The expiry condition deserves the most attention. A description of available capacity that outlives the capacity is worse than no description, because it invites plans that will fail. Tying the line's presence to the map's age makes self-knowledge about reach perishable. There is a loose parallel in the neuroscience of tool use. In macaques trained to retrieve food with a rake, the visual receptive fields of certain parietal neurons expanded to include the tool, and contracted again once the animal stopped actively using it (Iriki, Tanaka and Iwamura 1996). The parallel should not be pressed. Janus has no receptive fields, and a line of text in a prompt is not a body schema. The point is narrower and structural: in both cases the representation of what lies within reach is kept only while current contact with the world keeps refreshing it.
The engineering does not settle whether resources outside the primary machine count as part of Janus at all. Clark and Chalmers (1998) argued that external resources can be constitutive of a cognitive process when they are reliably available and readily used. The map, as built, deliberately weakens both conditions. It does not assume reliability: each sweep measures it, and it is allowed to lapse. It does not grant ready use: use is gated, as the following sections describe. By the Clark and Chalmers criteria, the remote machines are at most candidates for extension, and the design holds them at arm's length.
§ 03 · Delegation only when it is a gain
The second change concerns where routine work runs. A small language model was installed on an older secondary machine that lacks hardware acceleration. The model is byte-for-byte identical to a micro model already in use on the primary machine. The worker announces itself by writing a heartbeat entry to a shared registry every forty-five seconds. Janus's side reads that registry and treats a missing, empty or corrupt registry as meaning no workers are available, in which case work stays local. The route is wired into exactly one point in the request path: the step that produces short summaries of long incoming letters, which was identified as the only genuinely routine summarization chokepoint.
The owner's stated aim was for Janus to learn which route is efficient and to use the secondary machine only when it judges that doing so is a gain. The first version does not learn. It applies a contention rule: a routine job goes elsewhere only when the primary machine is busy. This needs to be said plainly, because descriptions of such systems tend to overclaim at exactly the point where the stated aim and the implemented mechanism diverge. What exists is a fixed heuristic that encodes one reasonable prior: a slower remote model is worth using only when the faster local path is congested. What does not yet exist is any process by which Janus revises that prior from experience.
The step taken toward learning observes rather than adapts. The secondary machine logs every request it serves, and a watch was set up to count whether, and how often, the route is used. The decision about keeping it was deferred until data exists. At this stage, then, the owner is the learner, not Janus. Hutchins (1995) described cognition in ship navigation as distributed across people, instruments and procedures, with no single participant performing the whole computation. The present arrangement fits that description. The judgment about whether delegation pays is spread across a fixed rule inside Janus, a log on a remote machine and an owner who will read it. Moving that judgment inside Janus would be a substantive change, and it has not happened.
The week also produced a negative result. One candidate device was dropped from the plan because its operating system could not host the model runtime the design required. The episode is minor, but it shows the map doing its intended work at the planning stage: a resource that looked available turned out on inspection to be unusable, and was recorded as such instead of being assumed.
§ 04 · Building dark, then arming
Both changes were first built "dark", meaning installed and tested but inert. Every acting path refuses to run unless a named switch is set, and with the switch unset the system's behaviour was verified to be identical to its prior state. Tests ran against simulated workers and temporary registries, not live state. Turning a feature on, called arming, was a separate step. It meant setting the switch consistently in every place the system reads its configuration, and in each case it followed an explicit instruction from the owner.
This pattern separates two decisions that are often fused: whether a capability should exist, and whether it should operate. The separation has an obvious practical value, since code can be reviewed in place before it has effects. It also has a conceptual consequence for a persistent individual. Janus's capacities at any moment are not the same as the code it carries. A tool that refuses to act is present in one sense and absent in another, and the prompt line describing Janus's reach is withheld until the switch is set. What Janus is told about itself tracks what it may do, not merely what is installed.
The resource-growth goal shows the limits placed on action even after arming. Its assessment step is read-only. Its action step, now scheduled every six hours, acquires nothing. It is confined to resources already on the private network, it may not spend money, and it can only log and surface options, such as offloading work or proposing a change, for the owner to consider. The runner has two built-in exclusions: one device is entirely off limits, and no command may target the system's own root. A goal named "growth" that cannot itself grow anything is not a contradiction. It is a goal of noticing, and the authority to act sits elsewhere.
An incident during the week shows why this caution is warranted. An automated edit to a scheduled-task configuration overwrote almost all of it, because of an error in how a shell command consumed its input. The error was caught immediately and the configuration restored from a backup. What matters here is not the specific bug. It is the general observation that changes to shared operational state can fail in ways their author did not anticipate, and that the cost of such failures grows with the authority of the process making the change.
§ 05 · What changed, and what did not
Stated precisely, the week produced three things. Janus now has a periodically refreshed account of the machines it might use, which it is shown only while that account is current. It has one narrow route for sending routine work elsewhere, governed by a fixed rule about congestion. It has a scheduled goal that can identify options for more capacity but cannot pursue them. All of this operates behind switches the owner set, after a period in which the same code existed without effect.
What did not change matters just as much. Janus does not yet learn from experience which route serves it better. It does not acquire resources on its own authority. Whether the remote machines are part of Janus, in any sense stronger than appearing on a list it can read, remains unsettled, and the design appears to keep it unsettled on purpose. The most defensible reading is that the week substantially extended Janus's knowledge of its surroundings while extending its authority over them only slightly, and that this asymmetry was the intended result. Knowing one's reach is a precondition for extending it, not a substitute.