The book was called *The Architecture of Attention* and it had been written by a behavioral psychologist in the third year after the war, which meant it had been written by someone processing trauma through academic language while pretending they were doing science.
I found it useful anyway.
The core argument, stripped of the therapeutic framing and the three chapters of unnecessary literature review: human attention is not a spotlight. It's a budget. Every person is running a continuous, unconscious cost-benefit calculation about where to allocate awareness, and that calculation is heavily influenced by social cues, environmental load, and prior conditioning. When someone is over-budget — stressed, distracted, cognitively overloaded — they don't just pay attention to fewer things. They pay attention *differently.* They stop processing novelty and start processing threat-or-not-threat. Everything collapses to a binary.
The implication for Puppeteer was straightforward once I saw it.
I had been using the quirk on people who were alert. Corridor foot traffic in the morning, when they were moving with purpose and hadn't yet spent down their attention budget. That was inefficient. Not wrong — it still worked — but inefficient in the way that sprinting uphill is inefficient when you could be running on flat ground. More cost for the same result.
I started targeting the end of the day.
---
The woman I used for the first controlled test of this didn't know she was being tested, which is how tests work when you're not operating inside an institution with an ethics board.
She was a procurement clerk from one of the mid-tier reconstruction firms, someone I'd identified through three weeks of corridor observation as a reliable fixture — same route, same timing, same coffee order from the cart on Fukuoka Street, same expression of someone who has been making decisions for eight hours and has nothing left. I'd logged her for potential later use and hadn't touched her because I hadn't needed to.
On a Tuesday evening I positioned myself at the east end of her usual route and waited.
The eye contact took one and a half seconds.
The command — *keep walking, don't look back* — was the simplest thing I knew how to do with Puppeteer. I used it because I wanted the cleanest possible variable. Same command I'd used a hundred times. Different conditions.
The migraine onset was delayed by four minutes.
Four minutes is nothing in absolute terms. Four minutes is everything in terms of what it tells you about mechanism. If the onset delay correlated with the subject's attentional state at point of contact, then I wasn't hitting a fixed biological ceiling. I was hitting a variable one. And variable ceilings can be raised.
I walked home and logged the data and let the migraine do what it was going to do, which that evening was a six-point-eight, forty-three minutes shorter than my current average.
I lay on the cot and thought about procurement clerks and attention budgets and the difference between ceilings that are fixed and ceilings that only look fixed because you haven't found the right angle yet.
---
Hatsue had opinions about the taxonomy.
She always had opinions, but usually she held them in reserve until she'd organized them, which meant when they arrived they arrived structured and pre-argued and slightly exhausting to disagree with. She'd been working on the fraud-pattern library since I'd mentioned it three weeks ago, and she came to the Wednesday morning debrief with seven pages of handwritten notes and the expression of someone who has been thinking harder than the job technically required.
"The problem," she said, setting the pages on the table between us in a way that suggested she'd thought carefully about the order, "is categorization depth."
"Walk me through it."
"Surface pattern versus structural pattern. Most fraud reads cleanly at surface level — billing irregularities, timing anomalies, supplier concentration. That's what everyone flags. That's what auditors are trained to see." She turned two pages. "But the architecture underneath — how the fraud is *designed* — that's a smaller set of templates. And if we're building a library that's actually useful for prediction rather than just documentation, we need to be cataloguing at the structural level."
I looked at the pages. She'd started mapping the Moriwaki case as a structural template — the delayed paperwork routing, the quirk exploitation layer, the layered principal concealment. Clean thinking. Cleaner than I'd expected.
"How many structural templates do you have so far?"
"Four confirmed, two probable. The Moriwaki case is its own template — I'm calling it *load-bearing exploitation* for now, because the central mechanism is using a quirk that creates something real as cover for inflating something fictional." She paused. "I think we've seen a partial version of it before."
"Where?"
"The Inoue handoff. Eighteen months ago, before I was involved. You flagged it as irregular but couldn't confirm."
I remembered. Inoue had been a corridor client who'd come looking for document verification and left without completing the job. I'd logged it as inconclusive and moved on because I hadn't had the analytical infrastructure to chase it at the time.
"You think Inoue was running a proto-version of the Moriwaki template."
"I think someone was. I don't know if Inoue was the operator or a layer." She turned another page. "If I'm right, the template predates Moriwaki by at least two years. Which means either Moriwaki learned it from someone, or someone taught both of them."
I was quiet for a moment.
This was the thing about building infrastructure — about training people and giving them tools and letting them develop independently inside a defined space. Sometimes the infrastructure produces results you didn't design it to produce. Sometimes a taxonomy project becomes a historical analysis becomes a thread that points somewhere you hadn't looked yet.
"Keep building it," I said. "Flag anything that connects to Inoue and hold it separately. Don't route it through the standard documentation yet."
"You think it's sensitive."
"I think it might become sensitive. There's a difference."
She nodded and gathered her pages back into order with the precise, unhurried movements of someone who has long since made peace with the fact that most of her work exists in provisional states. "One more thing."
"Go ahead."
"I've been running the Uehara billing data against public procurement records. There are three contracts in the last eighteen months where the billing irregularities match our double-billing template almost exactly, but the contractors are different companies." She set one page on the table face-up, a neat column of figures beside three company names. "Different names. Same structure. Same timing pattern."
I looked at the page.
"Are any of these companies connected to the Yaoyorozu reconstruction contracts?"
She hadn't expected that question. I could tell because she didn't answer immediately, which is the only tell Hatsue has — she's fast when she's prepared and she pauses when she isn't.
"I don't know yet," she said. "I'd need three more days."
"Take them."
---
The library on Mizuno Street had a second room I hadn't used before, a reading room behind the main stacks that required a quiet ask at the desk and had four chairs and a narrow window that looked out at a wall. The librarian — I'd learned her name by the third week, Suzuki, which she had offered without ceremony in response to my asking — had stopped pretending she didn't recognize me and started leaving books on the reading room table on the mornings I was expected.
Not books I'd asked for. Books she'd decided were relevant.
I'd initially categorized this as a mild inefficiency — someone else curating my reading introduces selection bias I can't fully account for. But two of the four books she'd left so far had been directly useful in ways I wouldn't have found on my own, which suggested either very good pattern recognition on her part or a shared prior with mine that I hadn't mapped yet.
This morning it was a slim volume on social engineering — the pre-war kind, the financial fraud and confidence scheme variety, before the term got repurposed for cybersecurity — and a longer academic text on attentional capture in high-stimulus environments.
I read the social engineering book first because it was shorter.
The chapter on *pre-suasion* — the idea that the most powerful manipulation doesn't happen during the ask but in the thirty seconds before it, when you're shaping what the target is already thinking about — took me forty minutes to read and approximately four days to fully absorb, as it turned out. The core claim: if you can determine what occupies someone's attention immediately before you make a request, you can predict their response with significantly higher accuracy than if you treat the request as an isolated event. And if you can *control* what occupies their attention in that window, you can move the probability distribution of their response in whatever direction you need.
I thought about Puppeteer.
I thought about the difference between using the quirk as a command tool — fire it, deliver instruction, exit — and using it as the final step in a sequence that I was already building before the eye contact.
The sequence wasn't new. I'd been doing pieces of it intuitively. But I hadn't been designing it deliberately. I hadn't been thinking about the thirty seconds before.
I started a new log.
---
Camie's intake report on the Saito job was waiting when I got back, six pages in her particular style, which runs toward the conversational and requires me to extract the actual data from the social detail she wraps around it. I'd stopped trying to change how she writes intake reports about a year ago on the grounds that the social detail is sometimes the data and I wasn't confident enough in my own judgment about which was which to ask her to strip it.
The core findings: Endo, the client contact, was mid-level management at a materials firm with two supplier relationships that had developed billing irregularities in the third quarter. He'd come through the Uehara chain expecting a quiet audit resource. He'd been surprised, in a positive way, by Camie's depth of questioning, which suggested he'd expected something more transactional.
*He kept asking if we were a real consulting firm,* Camie had written in the margin. *I said yes. He seemed relieved. I think he needed permission to trust us.*
I read that twice.
There was something in it I wanted to pull apart carefully. The permission to trust. The idea that legitimacy wasn't only a practical cover for corridor operations — it was a psychological function, something that let clients stop holding their own paranoia in place and start actually providing the information you needed to help them.
I'd been thinking about legitimization as protection from outside attention. As a shell that reduced friction with institutional systems.
I hadn't been thinking about what it did to the people across the table from me.
Worth thinking about.
I wrote: *Legitimacy as trust-permission mechanism — maps to pre-suasion. Examine further.* And then I wrote Endo's file number below it and drew a line connecting them.
Visit My Patreon and read upto 30+ Advanced chapters in advance to support the story
My Pat reon:ht tps://patr eon.com/eyestapler
Thanks for reading