The variable I changed was command complexity.
Simple commands had been my baseline from the start — *walk away, look elsewhere, give me that* — because simple was controllable and I was building a measurement system, and measurement systems require constants before they can afford variables. I'd been treating complexity as something to minimize. A source of error. The logical extension of the cost-reduction framework I'd built around the quirk.
I was wrong about that, in the specific way you're wrong when you've correctly identified a principle and then applied it one step past where it holds.
The book on attentional capture had a chapter on cognitive load theory that I'd read twice without fully absorbing, and then read a third time on a Wednesday morning in the Mizuno Street reading room while Suzuki reorganized something in the stacks with her usual quiet efficiency, and the third read produced something that felt less like understanding and more like recognition.
The claim: when a subject is processing a low-complexity instruction, the processing itself occupies a different neural substrate than when the instruction is complex. Simple instructions get routed fast, automatically, without real engagement. Complex instructions require the subject to construct meaning, which means they're *in* the instruction for longer, which means the window of influence is longer.
Puppeteer worked through eye contact. Eye contact opened the window. The command landed while the window was open.
If complex commands kept the window open longer on the subject's side — if the meaning-construction process was itself a form of attention that Puppeteer could ride — then the cost of complexity might be worth it. Not as a general principle. As a specific test.
I started building slightly more complex commands. Not much. Not *explain to me in detail why you first became interested in construction procurement.* More like: *go back the way you came and tell your colleague the meeting has moved.* A second object. A second action. Something that required the subject to hold two things at once.
The first three trials were failures in the sense that I couldn't tell what was happening differently. The data was noisy. The subjects were all wrong — mid-morning corridor traffic again, force of habit, exactly the wrong attentional state for what I was trying to measure.
I adjusted the protocol. End-of-day subjects only, same as the successful attention-budget test. Complex command structure. Lowest-stakes possible instruction. Log everything.
Trial four: a maintenance worker on his way out of a building I'd been monitoring. The command was *take the east stairwell instead of the elevator and check the third-floor water meter on your way out.* Specific. Plausible within his work context. Two embedded actions with a logical relationship.
The eye contact lasted two and a half seconds.
He took the east stairwell.
The migraine onset was delayed by nine minutes. Intensity: a five-point-six.
I stood on the street outside and felt something I didn't immediately have a word for. Not excitement — excitement is sloppy and I've been suspicious of it since the early corridor days when excitement was what made people take bad deals. Something more like the sensation of a calculation resolving. A variable clarifying.
The complexity was doing something. I didn't fully understand what. But it was doing something, and that was enough to continue.
---
Hatsue's three days produced seven pages and a question I hadn't anticipated.
We met on a Friday afternoon in the back room of the office, which I'd started thinking of as the working room because it had the better light and the table that was large enough to spread documents across without losing the edges. She'd brought printed copies of everything, which she always does because she doesn't trust digital-only record-keeping for active analysis. I'd come to agree with her about this. Paper is slower to access and faster to read, and during the part of the process where you're trying to see a shape rather than retrieve a fact, faster to read matters more.
"All three companies," she said, placing three sheets in a row, "have at least one contract where the billing irregularities match the double-billing template. All three are registered in the last four years, which is post-war incorporation, which tells us nothing on its own because most legitimate businesses incorporated or re-incorporated in that window." She placed a fourth sheet, perpendicular to the others. "But all three use the same incorporation agent."
I looked at the agent name.
"Beppuken Registrations," she said. "Mid-tier legal services, specializes in commercial incorporation and contract structuring. Has offices in three wards. Has provided incorporation services to forty-seven businesses in the last four years."
"How many of those forty-seven show irregularities?"
"I've checked nine so far. Four show the pattern. Two are inconclusive. Three are clean." She paused. "The clean ones are all in sectors with strong independent auditing. Healthcare supply chain. Pharmaceutical procurement. Places where the external audit pressure is high enough that the template is too risky to use."
"So the template-users are selecting for low-audit-pressure sectors."
"Consistently."
I sat back and looked at the four sheets and thought about what Hatsue had just described. It wasn't Moriwaki anymore. Moriwaki was one node. Beppuken Registrations was infrastructure. A shared toolkit that multiple operators were drawing from — possibly knowingly, possibly just because they'd been steered toward the same incorporation agent by someone who understood what that agent's network provided.
"Does anything in the Beppuken client list connect to the Yaoyorozu reconstruction contracts?"
Another pause. This one was different from the previous one — less *I haven't checked* and more *I was hoping you wouldn't ask that yet.*
"One company," she said. "Shinkawa Materials. Incorporated through Beppuken fourteen months ago. Currently holding a mid-tier materials supply contract on three Yaoyorozu-adjacent reconstruction projects." She placed a fifth sheet. "The contract value is modest. The billing is clean, so far as I can see. But the incorporation timing and the agent overlap are both present."
"And the projects are Yaoyorozu-adjacent, not Yaoyorozu-direct."
"She's not the contract holder. She's a design and load-assessment consultant on two of the three projects. Her name appears in the technical specifications as the primary structural authority."
Yaoyorozu Momo, whose quirk let her create materials from her own body, and whose reconstruction consulting work had been expanding in scope and contract value for the last year and a half. Whom I'd first flagged as a load-bearing element of the fraud architecture, back when I thought the architecture was only Moriwaki's.
The architecture was larger than Moriwaki. I'd known that in abstract terms for a while. This was the first time I was seeing the edges of how much larger.
"Don't touch Shinkawa yet," I said. "Log it and hold it and keep mapping the Beppuken network. I want to know how many of the forty-seven before you stop."
"All forty-seven?"
"All forty-seven."
She wrote a note without visible reaction to the scope of the work. This is one of the things that makes Hatsue invaluable in a way that's difficult to quantify on a job log — she doesn't treat the scale of a task as relevant information about whether she'll do it. She treats it as relevant information about how long it will take.
"Two weeks," she said. "Maybe three."
"Take three. Do it right."
---
The Saito job resolved on a Thursday.
Not dramatically. The billing irregularities were what Endo had suspected — double-billing from a mid-tier supplier who had correctly identified that his contract didn't include a mandatory independent audit clause, and had been running a modest but consistent inflation on materials invoices for eleven months. Clean template. Well-documented. Hatsue's taxonomy had made the identification faster than it would have been before; we'd matched the pattern within four hours of receiving the full billing records and spent the remaining time building documentation clean enough to take to an external auditor without raising questions about how we'd worked it.
The fee was appropriate for the work. Not exceptional. The work hadn't been exceptional — it had been competent, which is its own category and shouldn't be confused with the other one.
What was interesting was the debrief meeting with Endo.
I ran it myself, which I hadn't originally planned. Camie had done the intake and the initial relationship management, and the logical flow was for her to handle the close as well, maintain continuity, keep the client experience coherent. I changed the plan forty-eight hours before the meeting because Camie's intake report had flagged something I wanted to see directly.
*He kept asking if we were a real consulting firm. I said yes. He seemed relieved. I think he needed permission to trust us.*
I wanted to understand what Endo looked like after the permission to trust had been granted and the work had resolved well. What does a client look like when they've come from a paranoid approach posture and had the paranoia resolved by results?
The answer, as it turned out, was: unusually forthcoming.
We sat in a neutral meeting space, the kind of short-term office rental that exists in Musutafu for people who need a table and a door that closes and no other implications. Endo was a compact man in his mid-forties with the particular wariness of someone who has spent years working adjacent to systems he doesn't fully understand and has developed a finely tuned sense for when something is wrong even when he can't articulate why.
That sense had brought him to us. The fact that we'd confirmed and documented what his instincts had told him was happening seemed to have recalibrated something in how he was sitting.
"I expected it to be slower," he said, which wasn't the first thing I'd expected him to say.
"Slower how?"
"The identification. I thought you'd need more time to work through the supplier chain. We've had two internal audits in the last year and neither of them caught it." He looked at the documentation I'd placed on the table. "How long did it take you?"
I gave him the truthful answer because the truthful answer was more useful than a comfortable one. "Four hours of active analysis after we had the full records. The pattern matched a known template."
"A known template."
"We catalogue fraud structures. When you see enough of them, they resolve faster."
He was quiet for a moment in the way people are quiet when they're recalculating something. "How many of these do you see?"
"Enough to have built a catalogue."
Another silence. Then: "The supplier — Tanaka Materials — I've been working with them for three years. Two of those years were clean." He paused. "When did it start?"
"Eleven months ago. The change correlates with a shift in their internal accounts management. Someone new came in at the mid-level and changed the invoicing workflow."
"So it wasn't Tanaka himself."
"Almost certainly not. The pattern is consistent with a rogue actor inside the supplier's operation who identified a gap in your audit structure. Tanaka may not know it's been happening."
Endo absorbed this. "That makes it more complicated."
"Yes."
"I've been trying to decide whether to report it or absorb the cost and adjust the contract structure going forward." He looked at the documentation again. "If it's a rogue actor inside Tanaka's operation, that changes the calculus."
"It does," I said. "It also means Tanaka has a liability exposure he's not aware of. Which gives you a different kind of leverage than you'd have if the fraud were authorized at the principal level."
He looked at me.
"I'm not recommending an approach," I said. "I'm noting the structure."
"No." He almost smiled. "But you're very good at noting structure in ways that make the options clear."
I said nothing.
"The referral I came through — Uehara's people." He straightened slightly. "They said you were thorough. They didn't say you were fast."
"Thoroughness and speed aren't usually in competition. It's a question of whether your tools match the problem."
"And your tools matched this one."
"Our tools matched this one."
He looked at me for a moment longer than the conversation required. Not the way people look when they're trying to read you. The way they look when they've already read you and they're deciding what to do with it. "If I come across something more complex," he said, "I'd like to come back."
"We're generally available for referrals at this tier."
He nodded and stood and picked up the documentation, and I watched him leave with the particular gait of someone whose posture has changed since they sat down — a fraction less compressed, the weight distributed differently.
*Permission to trust, granted, confirmed, and now transferred forward.*
He would refer someone. Within two months, probably.
I noted the prediction in my log on the walk back to the office. The corridor was busy in the late afternoon the way it gets when the reconstruction crews change shifts, and I moved through it with the particular attention management I'd been practicing — aware of everything, reactive to nothing, the baseline hum of the city as information rather than noise.
Six blocks from the office I stopped at the corner of Hayashi and Mitsu and looked across the street at a building I'd been tracking for three weeks.
Yaoyorozu Structural Consulting. Second floor, the whole south face of the building. The lights were on.
She worked late. I'd established that in the first week. The pattern was consistent — she arrived early and she stayed until the work was done, which suggested either very high internal drive or project loads that exceeded normal capacity. The reconstruction contracts were generating more demand than the operation could absorb at current scale, which meant either she was scaling up or she was becoming a pressure point.
I looked at the lights for thirty seconds.
Then I turned and walked the remaining six blocks to the office and sat down at my desk and opened the Beppuken file and added *Shinkawa Materials / Yaoyorozu-adjacent* to the active monitoring list with a priority flag.
Not yet. The shape wasn't clear enough yet. Moving before the shape is clear is the mistake I've watched other corridor operators make repeatedly — they see a thread and they pull it before they've mapped what the thread is attached to, and then the thing they pull collapses or pulls back and they're exposed without cover.
I was not going to make that mistake.
I was patient.
I wrote: *Lights on at 21:40. Pattern consistent. Possible capacity stress. Monitor for contract volume change and personnel addition.*
Then I closed the file and opened the training log.
Five-point-six. Best single result so far.
The ceiling was stepping.