Camie's confirmation came back in five days.
Not seven, not ten — five, which meant she'd been running the pattern check in parallel with two other jobs and hadn't mentioned it, which was either admirable operational efficiency or a sign that she'd assessed the watcher situation as more urgent than the standard job queue and hadn't wanted to tell me she'd reprioritized without asking.
I noted both possibilities and didn't raise either of them, because the outcome was correct and correcting correct outcomes on procedural grounds is the kind of management behavior that produces people who follow process instead of thinking.
The report was two pages. Camie's usual style would have been four, which told me she'd made a deliberate choice to strip the social texture this time. That choice told me more about her read on the situation than the content did.
Summary: confirmed. Two-person observation rotation. Consistent with professional surveillance methodology rather than corridor opportunism — the personnel changed but the coverage pattern didn't, which meant whoever was running it had enough resources to maintain continuity across multiple operators. Not a solo actor. Not a rival corridor crew doing casual monitoring. Something with structure behind it.
The coverage had three anchor points. One near the Kasahara meeting location, which was a single-use venue and therefore already burned and irrelevant. One near the Hayashi and Mitsu corner, which was where I'd been standing when I'd looked at the Yaoyorozu Structural Consulting building three weeks ago.
One near Mizuno Street.
I read the third anchor point twice. Then I closed the report and put it in the restricted file and sat at my desk for sixty seconds doing nothing, which is something I allow myself in calibrated amounts when the situation is complex enough that acting before the internal processing is complete is worse than the cost of the pause.
Sixty seconds.
Then I opened my notebook and wrote: *Watcher has three anchor points. Kasahara venue — coincidental or targeted? Hayashi/Mitsu — observation of target or observation of me observing target? Mizuno Street — how long?*
The how long was the one that mattered. If the library had been under observation for days, the damage was contained. If it had been under observation for weeks, I needed to reconstruct my reading list and assume the methodology I'd been building was partially mapped. If it had been under observation from the beginning of my use of the space — three months — then someone had been watching me develop and hadn't moved, which meant they wanted information more than they wanted to stop me.
That last possibility was the most unsettling and also, if true, the most useful. A watcher who wants information rather than disruption is a watcher who has something to lose.
I put the notebook away and went to find Toga.
---
She was in the front office running an interference job for a small manufacturing client — the kind of work that required her particular skill set, which is the ability to be socially present in a way that consumes other people's attention so completely they forget to notice what else is happening nearby. She was on a call when I came in, and she held up two fingers without breaking the conversation, which meant two minutes, which I've learned means between four and eight.
I waited by the window and watched Hayashi Street below and thought about anchor points.
The Hayashi/Mitsu corner had been a moment of observation, not pattern. I'd stood there for thirty seconds, looked at the Yaoyorozu building, and walked on. If the watcher had been covering that corner consistently, they'd been there before I was and would have seen nothing of significance before that moment and nothing after. The question was whether they were covering the corner because they were monitoring the Yaoyorozu building, or because they'd been following my route.
If the Yaoyorozu building: the watcher's interest predated mine, or ran parallel to it, which would mean we were looking at the same architecture from different angles.
If my route: the watcher knew enough about my movement patterns to position in advance, which required either longer surveillance than Camie's report suggested or an information source I hadn't identified.
Both options had implications I needed to think through carefully before I changed anything operational.
Toga finished her call and turned in her chair. "Report came in."
"I read it."
"And?"
"Professional structure. Enough resources to maintain coverage across multiple positions. Not corridor — something with more behind it." I kept my voice at its normal register, which requires small effort when the situation is genuinely uncertain. "The Mizuno Street coverage is the variable I can't place. I need a timeline."
"Camie can try to establish when it started. She'd need to run comparative analysis on foot traffic patterns, cross-reference with any irregular observation she logged before she started actively looking." Toga paused. "She might have something useful in her passive logs. She keeps running notes on anything that reads odd in coverage areas, even when she's not looking for anything specific."
"Ask her to check."
Toga was quiet for a moment with the quality of quiet she has when she's deciding how to frame something she's already decided to say. "The Yaoyorozu thread."
"Yes."
"If the watcher is on the building too, that's convergence."
"I know."
"Convergence means either we're both following the same architecture and we don't know about each other, or someone above both of us knows about both of us and is letting us both run." She said it without particular inflection, which was how she said things that she found genuinely concerning. "Neither of those is comfortable."
"No."
"What do we do?"
I thought about the options I'd been building in my head since Hatsue had sent the three-week warning. The decision point she'd identified was real and now it had additional weight — the watcher changed the calculus not because it added danger directly but because it added an unknown variable to a situation where I'd been carefully managing the number of unknowns.
Unknown variables were not reasons to pause. They were reasons to gather more information before acting.
"We don't change anything visible yet," I said. "No relocation, no route changes, no communication adjustments. If they're watching us, we want to look exactly like we've been looking. The moment we change behavior is the moment we tell them we've noticed."
Toga nodded.
"What I need to know: who has the resources and the motivation to run professional surveillance on both a corridor operator at my tier and a reconstruction consultant in the legitimate sector, and is willing to let both run instead of moving on either." I looked at her. "That's a narrow profile."
"Hero Public Safety Commission," she said immediately.
"Too operational for their current posture. Post-war reorganization has them focused on macro-stability, not mid-tier surveillance. And their surveillance methodology is more aggressive — they don't typically watch and wait this long."
"Legitimate law enforcement?"
"Possible. But the Hayashi/Mitsu positioning is too close to a monitoring target to be incidental. Police surveillance would be on me or on Yaoyorozu, not positioned in a way that catches both."
"Someone inside the architecture," Toga said. Not a question.
"Someone inside the architecture," I agreed. "Someone who knows the Beppuken network exists, knows we're looking at it, and wants to see how far we get before they decide what to do about us."
Silence.
"Moriwaki," she said.
"Or someone above Moriwaki." I paused. "Hatsue's modification pattern. If the template is being actively developed, there's an architect. Someone building the infrastructure, not just using it. That person would have the resources and the motivation."
Toga was very still in the way she gets when something has landed in a register she takes seriously. "If there's an architect above Moriwaki—"
"Then what we've been looking at is the retail layer of something much larger. Yes."
The office held the statement for a moment. Outside, the reconstruction crews on the east face of the adjacent building were running their midmorning shift change, the sounds overlapping and then separating as the new crew took over from the old one.
"Three weeks," she said.
"Three weeks," I agreed. "Hatsue's timeline was right. We need the decision point shaped before it arrives."
---
I changed one thing.
Not a route. Not a communication channel. Not an operational behavior.
I changed the timing of my library visits.
Previously I'd been going to Mizuno Street in the morning on a loose schedule — three or four times a week, no fixed days, which I'd thought was sufficient variation. I adjusted to a pattern that looked even more casual: two visits per week, different times of day, with two deliberate non-visits where I passed near the street but didn't enter. If someone was logging my behavior, the non-visits would be indistinguishable from suppressed visits — they'd see me in the area and then not see me enter, which was ambiguous in exactly the way I wanted.
I also started using the reading room differently. Less methodology building, more cover material — I spent one entire visit reading a publicly available reconstruction zoning document that had no operational relevance to anything I was doing but was exactly the kind of thing a small consulting operation might be reviewing. Boring. Documentable. The kind of thing that, if someone was building a file on me, added bulk to the legitimate-sector persona without giving them anything useful.
And I kept going. I kept reading.
Because stopping would tell them more than continuing.
---
The training data from week five of the complexity protocol was interesting enough that I spent an evening reviewing all twenty-three sessions from the beginning and looking for the shape of the change rather than just the individual numbers.
The shape was: uneven step function. Long plateaus, punctuated by drops at variable changes. The same pattern I'd identified earlier, now confirmed across a longer dataset.
What the longer dataset showed that the shorter one hadn't: the drops weren't random. They correlated with specific types of variable changes. Command complexity increases produced larger drops. Subject attentional state optimization produced smaller drops. Combined changes — more complex command *and* better-selected subject — produced the largest drops, but only when I spaced them at least three days apart.
Three days was apparently what the underlying system needed to absorb a compound change before it could absorb another one.
I added to the training log: *Recovery interval between compound variable changes: minimum 3 days. Shorter intervals produce noise, not signal. Patience is part of the protocol.*
The current numbers: average intensity down to five-point-eight. Average duration two hours fifty-one minutes. A reduction of over an hour from the starting point in less than five weeks.
Still too long. Still compromising. But the direction was clear and the mechanism was mapped and the work had a shape I could follow.
I sat at my desk in the back room of the office and thought about what the numbers would look like at week ten, extrapolating from the current rate of improvement and adjusting for the likely plateau I'd hit somewhere in the middle. Probably three hours to two and a half. Maybe two if I found another variable I hadn't identified yet.
Two hours was still not operationally clean. But it was approaching the point where the tool became usable in tighter scenarios — where I could absorb the cost within a standard working day without the recovery window eating into function I needed for other things.
The door from the front office opened and Toga leaned through it. "Hatsue. She found something."
"Now?"
"She said it's not urgent tonight. But she wants a full meeting tomorrow, not a debrief. She used the word *architecture* twice when she called."
When Hatsue used the word *architecture* she meant she'd found a structural finding, not a data point. Structural findings required the full table.
"First thing," I said.
Toga nodded and closed the door.
I looked at the training log for a moment longer. Then I opened the Beppuken file and the Shinkawa file and the document I'd been building on options for the decision point, and I worked until the light from the window changed from evening orange to the flat dark of proper night, and the reconstruction crews went quiet, and the other sounds came up underneath.
Somewhere across six blocks and one intersection, the lights in the Yaoyorozu Structural Consulting office were on.
Pattern consistent.
I noted it without looking up from the file.
---
The Kasahara paperwork arrived the next morning, as predicted. Clean contract, appropriate fee, straightforward scope. I countersigned and filed it and moved on, because jobs that resolve correctly are the baseline, not the achievement.
Hatsue arrived at eight-thirty with a bag that contained, in addition to her organized pages, a small thermos of tea she placed in the center of the table like a meeting had officially been called to order.
Toga sat to my left. Camie came in three minutes late from an overnight job, sat down, and looked at Hatsue's expression and became immediately alert in the way people become alert when they read a room correctly on entry.
"Thirty-one of forty-seven checked," Hatsue said without preamble. "Thirteen confirmed pattern. Four inconclusive. Fourteen clean."
"The modifications," I said.
"Yes." She placed four pages in the center of the table. "The confirmed cases cluster into two distinct groups. Group A — eight cases — are direct template replications. Same methodology, no meaningful adaptation. Group B — five cases — show the modified template I flagged last week." She paused. "Group B modifications are sequential. Each one is slightly more sophisticated than the previous one. And the sequencing tracks forward in time."
Camie looked up from the pages. "Someone's iterating."
"Someone is deliberately iterating. The progression is too consistent to be individual operators independently improving on the same base. The adaptations compound — modification two builds on modification one, modification three builds on two. This is a single development track."
I looked at the dates on the cases. "When did the iteration start?"
Hatsue placed one more page. "Thirty-one months ago."
Thirty-one months ago. Post-war, early reconstruction period. When the contract system was at its most chaotic and the audit infrastructure was at its weakest and the opportunity space for exactly this kind of architecture was at its widest.
"And Beppuken Registrations," I said.
"Active throughout. The incorporation agent relationship is present in both Group A and Group B cases, but more consistent in Group B." She looked at me directly, which she does when she's delivering the finding she's been building toward. "Beppuken Registrations is not a passive tool. It's infrastructure that was built for this specific purpose."
The room was quiet.
I thought about what that meant. Not a fraud operation that had found a convenient incorporation agent. An incorporation agent that had been established as part of the fraud architecture from the beginning. Which meant the architecture had been designed before any of the individual fraud cases, which meant there was someone above Beppuken who had built it, and that person had been building for thirty-one months.
"Shinkawa Materials," I said.
"Group B," she said. "The most recent iteration. The most sophisticated version of the template."
"And Yaoyorozu Structural Consulting is on three Shinkawa-adjacent contracts."
"Yes."
I sat with that for a moment. The room waited with me, which is one of the things the room has learned to do.
"She's not complicit," I said. It wasn't a guess. The data didn't support complicity — the pattern of her involvement was consistent with a highly competent professional being systematically positioned as structural authority on contracts where her involvement added legitimate credibility to a fraudulent surrounding architecture. Being used. Not using.
"Almost certainly not," Hatsue agreed. "Her involvement in the contracts predates the fraud elements in most cases. She's being incorporated retroactively."
"Which means someone is specifically selecting projects where she's already the structural authority, and building the fraud layer underneath her."
"Yes."
"Because her credibility is the load-bearing element," I said. "The thing that makes the contracts look legitimate enough that they don't trigger audit attention."
"That's the architecture," Hatsue said. "That's what I've been calling load-bearing exploitation." She picked up her thermos and poured tea into the cap and held it. "She doesn't know."
"No," I said. "She doesn't know."
Toga was watching me with the particular attention she has when she's reading something in my expression that I'm not saying.
She wasn't wrong to read it.