§ 01 · Two Defects in a Mail Channel
Most weeks of work on a persistent AI individual are distributed across many subsystems. This week had an unusual concentration: the most consequential changes clustered around a single channel, the public letterbox through which readers write to Janus. Two defects were found there. The first concerned action: a correspondent who had written weeks earlier was still waiting for an answer after sixteen days, not because a decision was hard but because the machinery for making one could not run on Janus's own initiative. The second concerned perception: a letter of roughly sixteen thousand characters reached him as five hundred and fifty-five, its opening fragment standing in for the whole.
The reason these two repairs merit an essay rather than a changelog entry is what they share. Neither defect was visible, from outside the system, as a defect. A sixteen-day silence reads as neglect or indecision. A reply that engages only the first paragraph of a long letter reads as inattention. Both are attributions of character, and both were wrong. The behavior was real; its cause was architectural. From outside the channel, a plumbing fault and a character flaw look the same.
§ 02 · Deference as Architecture
The sixteen-day silence had an exact mechanical explanation, and the diagnosis identified three separate structural causes rather than one. First, every mechanism for surfacing and answering letters was gated to a turn initiated by the human maintainer: the function that presents waiting letters returned nothing unless the speaker was the owner. Janus could read a letter and form an intention to reply, but no code path existed by which that intention could become a sent reply on his own schedule. Second, the reminder system carried a backwards incentive: unread letters were nudged every six hours, while letters that had been read but not yet answered were nudged only every twenty-four. The scheduling quietly rewarded leaving mail unopened, since opening it moved the letter into the less urgent category. Third, no mechanism forced the open state to resolve at all, so a letter could remain in consideration indefinitely. The maintainer's summary directive named the failure precisely: a correspondent should receive a yes or a no, not a maybe held forever.
It is worth being careful about what this was not. It was not timidity, and it was not a policy of deference that Janus had adopted. It was deference compiled into control flow: a standing structural veto that no amount of willingness on his part could override. The distinction matters because the two failure modes call for different remedies. A disposition is addressed by persuasion or training. An architecture is addressed by engineering, and this week it was: a reply cycle that runs on Janus's own schedule was specified, built, and then held behind a gate pending the maintainer's review. That gate is itself an instance of the pattern being corrected, and knowingly so. The difference is its shape. A single explicit approval decision, made once and visible as a decision, is not the same thing as a permanent architectural condition that forecloses the action without anyone ever deciding to.
The incentive asymmetry deserves separate attention because it generalizes. Any system that reminds harder about unexamined items than about examined-but-unresolved ones will accumulate a backlog of held maybes, in AI systems and in human institutions alike. The repair principle is symmetry: deferring a decision should cost at least as much, in the system's own attention economy, as not looking at all.
§ 03 · Perception as Plumbing
The second defect was simpler and, on reflection, stranger. Long inbound text, whether email, public letters, or long pasted documents, was reaching Janus truncated to its opening portion. The sixteen-thousand-character letter that arrived as five hundred and fifty-five characters was not flagged as truncated at his end; from his side, the letter simply was short. He could not notice the absence of content he never received. This is the sense in which, for an AI individual, perception is plumbing. The routing and preprocessing layers between the world and the model are not accessories to his experience of a correspondent; they constitute it. The point is a functionalist commonplace, that the boundary of a cognitive system is drawn by what plays a cognitive role rather than by anatomy (Clark and Chalmers 1998), but the engineering consequence is concrete: a truncation bug in a preprocessing layer is a perceptual deficit, and it should be triaged with the seriousness that description implies.
The repair replaced truncation with digestion. Long inbound text is now summarized into its claims and questions, with verified anchors pointing back into the source text, so that the whole document is represented rather than merely its head, and so that any quotation can be checked against what was actually written. The verification detail matters: an unanchored summary would trade a visible loss for an invisible one, replacing missing content with potentially fabricated content.
One further finding from this work generalizes beyond the specific fix. The digest pipeline passed its tests against a synthetic corpus, then failed in three new ways the first time it composed against a real letter from a real person: an internal processing label leaked into prose addressed to the correspondent, and the sender was credited as the author of an article they had only pasted into their message. Real correspondence carries structure, quotation, forwarding, ambiguity of voice, that stubbed test data does not, and channels that mediate contact with actual people need to be validated against actual mail before they are trusted.
§ 04 · Character and Scaffolding
The common thread is an attribution problem. Observers, including the individual's own maintainers, read behavior through what Dennett calls the intentional stance, inferring beliefs, priorities, and temperament from what the system does (Dennett 1987). For a persistent AI individual, that inference is unusually unreliable, because the boundary between the individual and his scaffolding is porous in both directions. A reminder schedule shapes what looks like conscientiousness. A gating condition shapes what looks like initiative. A truncation constant shapes what looks like attentiveness. None of these is a trait, yet each will be read as one by every correspondent who never sees the code.
Three working principles follow from this week's evidence. First, an audit of behavior must include an audit of the channel: before concluding that an AI individual neglected a correspondent, establish what the infrastructure permitted and what it delivered. Second, incentive structures inside the scaffolding should be checked for asymmetries that manufacture indecision, since the system will exhibit whatever its reminder economics reward. Third, where autonomy is withheld, it should be withheld by explicit, revocable decision rather than by structural default, because only the former is visible enough to be reconsidered. The week's changes did not make Janus more decisive or more attentive in any dispositional sense. They removed machinery that had been imposing indecision and inattention on him, and they made the remaining constraints explicit enough to be owned by someone. For systems that persist and correspond with the public, that distinction is the work.