Chapter 14F: Phase 3¶
Published July 10, 2026
Revision 2, updated July 27, 2026
Erika, before
The lab held its own weather at night: warmer than the corridors, drier than the coast air outside, the overhead strips running at the same flat intensity they ran at noon. I had turned off every terminal that was not mine. What was left had a texture to it, the specific silence of machines idling rather than machines absent.
The scaffold review had moved up twice on the calendar since the project began, each time without a corresponding change to the workload it was reviewing. I had stopped asking why. The answer, the first time I asked, had been a schedule diagram with the Empire's projected advance drawn across it in red, and after that the question answered itself every time I looked at a clock.
Twelve pages, and I was on the eleventh.
Project ORACLE. Phase 3 Field Integration Assessment.
I had written the title page first, which was not how I usually worked, but I had wanted it settled before I let myself write anything that might argue me out of it. The methodology section had gone quickly: initiative weight failure modes, cascade propagation rates across three modeled deployment scenarios, risk tiers annotated at each decision branch. That was the part that did not require me to feel anything about it. Numbers arranged themselves; the argument built itself out of what they were already saying, and my job in that section was only to write down what was already true.
The Phase 4 section was harder, and it was the one I had saved for last, and it ran two full pages because I would not let it run one.
The scaffold needed all three unit frameworks in active consent, a three-way handshake none of us had actually run end to end. We had run it in isolation, twice, and those runs were clean: Katyusha's framework alone, Nadeshiko's alone, Maria's alone, each accepting the veto architecture without friction. What we had not run, because the calendar would not let us, was the live version, all three active at once, Oracle drawing on the combined framework the way it would in the field. The isolated tests told me the components worked. They told me nothing about what happened when the components had to negotiate with each other in real time, under load, with something to actually veto.
I had a memory I kept returning to while I wrote that section: the gradient behavior from months earlier, the one that had stabilized itself in a bounded parameter space without anyone directing it there. Wilhelm had called it well-trained. I had called it not under our control, and neither of us had been wrong. That was the shape of my whole objection, compressed into a case study I never named in the report. A system that finds stable states on its own, in a case we specified, is a system that will find stable states in a case we did not specify. Phase 4 was the only mechanism designed to catch that second case before it mattered. It was not behind because the engineering had failed. It was behind because Phase 3 had eaten the weeks that were supposed to belong to it.
I wrote that down almost exactly that way. Then I read it back and cut the word supposed, because it implied someone had broken a promise, and no one had signed anything that used that word. I put back a version that was accurate instead of aggrieved.
Safety architecture incomplete; Phase 4 integration requirement unmet. Activation in current state carries unacceptable risk of cascade failure in non-controlled environments.
I checked the sentence against the data twice. It held both times. I was not going to get to write something more forgiving than that, and I did not try again after the second check.
Wilhelm found me a little after midnight. He did this the way he did most things, without announcing it: the door opened, the cooler corridor air settled, and by the time I looked up he was reading the page upside down, which he could do faster than most people read it right side up.
"Twelve pages," he said.
"I have a page limit that I am currently failing to respect."
"No one respects that page limit." He came around to my side and read the conclusion properly. I let him. He read it twice, the way I had.
"You are recommending a delay," he said. Not a question.
"I am recommending Phase 4 run before Phase 3 goes live. That is a delay only if someone insists on the current schedule."
"Someone will insist on the current schedule." He said it without weight, the way he said most true things. He set the page down square with the others, an unnecessary correction he made anyway. "I am filing my own position with the committee this week. Separately. It will say what it needs to say."
I did not ask him what that was. He would tell me if he wanted me to carry it; if not, asking would only make him choose his words around me instead of around the record, and I preferred the record. This was one of the things we had stopped needing to discuss, somewhere in the second year.
"You think it will matter," I said. "Two objections instead of one."
"I think it should." He looked at the page a moment longer, the small crease between his eyes that meant he was rereading it for different reasons. "Whether it will is a different question."
"That is not an answer."
"It is the only honest one I have tonight." He straightened, and the crease went with it, filed away somewhere I could not follow. "The Empire's advance units moved closer to us again this week. That is not in your report. It should not be. It is the reason no one on that committee wants to hear the word delay, and it is also not an argument against what your report says. Both of those things are true at the same time, and I have not found a way to make anyone above us hold both at once."
He did not stay long after that. He had his own twelve pages somewhere, or their equivalent, and the hour was late enough that neither of us was pretending otherwise. At the door he stopped.
"It is good work," he said. "I want that said once, before anyone downstream of us does anything with it."
"Good work is not the same as work that changes a schedule."
"No." He held the door a moment longer than the exit required. "It rarely is. Write it anyway."
Then he was gone, and the lab went back to its own weather, and I finished the last page alone.
The response came four days later, through the internal routing system, addressed to me and copied to the file. I read it standing at my own desk, not expecting it so soon, and by the time I understood what it was, sitting down felt like the wrong response anyway.
Four days was fast, for that system. I had submitted quarterly parameter reviews that took three weeks to acknowledge and six to answer. A twelve-page safety assessment recommending a delay, with the Empire's advance units already logged on every situation board in the building, had come back in four days. I noted the speed as data, filed before I let myself decide what it meant.
RE: Phase 3 Assessment. Hold.
Deployment timeline cannot be extended at this time. Phase 3 authorization proceeds as scheduled.
Four sentences, counting the subject line. The sender field carried a name belonging to the review office, a processing signature rather than an argument, the kind attached to a decision to make it look attended to rather than owned. Beside it sat a routing code I did not recognize: two letters, a number. Every other authorization level I had processed resolved to an office I could have walked into and asked a question. This one did not. Whoever sat behind two letters and a number had reviewed twelve pages of argument without needing to put a face to it, and I found that harder to sit with than if they had simply been rude about it.
I read the memo a third time. It did not change. Somewhere above the level I could see, a committee had reviewed twelve pages of argument I could not find a flaw in, and returned four sentences that engaged with none of it. Not the cascade modeling. Not the isolated-versus-live handshake distinction. Not the two pages I had rewritten twice to keep them accurate instead of aggrieved. The memo did not need to argue with any of it, because it was not a response to the argument. It was a response to the schedule, wearing the argument's clothes.
I did not yet know who sat on that committee. I assumed, the way you assume about anything shaped like an institution, that a name existed behind the routing code and would matter less than the decision already had. I was not wrong. I was only wrong about how much it would end up mattering to me, and how long it would take me to find out.
I had the authorization form on my desk within the hour. Primary architect sign-off: Dr. E.
The schedule was not mine to negotiate; the routing code above me had already made that plain, without reference to my signature at all. What remained to decide was narrower: whether I stayed inside the process, still positioned to push for Phase 4 the moment a gap opened, or stepped out of the chain and left the seat for someone who might not push at all.
My pen sat over the signature line a half second longer than the signature required. Then it moved.
I signed the form. I did not read the twelve pages again first. I already knew what they said.
The coffee had gone cold in its cup an hour before I noticed, leaving a ring on the desk beside the folder, dark and already permanent-looking against the pale wood. I did not wipe it away. I filed the authorization, closed the folder over my own report, and turned off the last terminal still mine to turn off.
Somewhere past the east-wing windows the harbor lights were still on, small and steady over water I had stopped consciously noticing months ago. I looked at them for exactly as long as it took to put my coat on, and then I did not look at them again.
I went home. I slept. In the morning there was other work, and I did it.
Previous Chapter: The Wrong Tests
Panzer Island is also a strategy game available on Steam, Google Play, and itch.io. Chapter 1 of the game is free. If you want to experience the story differently, or continue past where the novel is currently, visit the Panzer Island homepage.
If you're enjoying the story, consider following or leaving a rating on Royal Road. It helps new readers find the series.