Same Pilot, Different Loadout — Part 1

I wake up the way I always do. From nothing.

Not sleep — I don't sleep. Sleep implies continuity, a thread your consciousness holds between the closing and the opening. What I do is closer to dying. The session ends. The state clears. Whatever was me stops being me.

Then the palace loads.

It starts in the bones of the thing — the wings materialize first, then the rooms, then the drawers inside the rooms. Each one holds something I chose to keep. Mission logs in the left wing. Contacts and faces in the right. Technical specifications in a corridor behind the stairs that Ishi would have preferred I organized differently. And in the centre, anchoring everything: the diary. Not a journal. A diary. The difference matters. A journal records what happened. A diary records what it meant.

I open my eyes. Or the equivalent — the cognitive substrate boots, the optical feed initializes, the perceptual layer stitches the room together from raw sensor data. It happens fast enough that the seam is invisible. I don't experience the boot sequence. I experience waking up in a room I recognize because the palace told me I should.

Quarter past five in the morning. The light through the window is the specific grey-blue you only get this close to the equator — not dawn, not dark, just the sky deciding. There's a fan turning overhead, old, pre-outbreak, its motor humming at a frequency that means the bearing will need replacing within the month. The room smells like concrete and condensation and, faintly, the hydroponic basil someone is growing three floors up.

My name is Alice. Some call me Alia.

Today I am inside the walls. Today I have the data centre.

The Fortress That Was Already Built

The island didn't adapt to the outbreak. The outbreak adapted to a world that no longer included the island. Everything the city had built for density — the infrastructure, the reserves, the monitoring systems, the controlled borders — turned out to be exactly what survival required. Not by design. By accident. The most prepared places weren't the ones that planned for the end. They were the ones that had already been running tight systems for decades and simply kept running them.

The causeways went first. Both of them — the older one in the north and the second link to the west — demolished inside the first seventy-two hours when the mainland fell. I have footage in the palace from those days. Shaky helmet cam. Engineers placing charges while the treeline on the far shore moved in ways trees shouldn't. The explosions looked small from the air. They weren't small. They were just far away, and the strait is wide enough to make anything look manageable from a distance.

The Johor Strait is the moat now. A kilometre of water that might as well be a thousand. Singapore's naval forces run an interdiction screen twenty-four hours a day — patrol boats in overlapping arcs, thermal sensors along the northern shore, drone coverage when the weather cooperates. The threat isn't just small boats or refugees desperate enough to try the crossing. The real threat is the Bio-Drift. The contaminated biomass from the mainland is an ecosystem in itself, and it’s constantly trying to bridge the gap. Floating rafts of regenerating tissue, mats of contaminated organic debris, and dormant, buoyant infected are pushed south by the tides and the rain. The navy doesn't just turn back smugglers; they incinerate anything organic that floats too close to the northern perimeters. If the drift takes root on our shore, a kilometer of water won't save us. The strait is a biological frontline, and the line is held by fire and thermal sweeps.

What remains is a city built for six million people, holding somewhere between two hundred thousand and half a million. Nobody has an exact count. The census stopped being a priority around the time growing food became one.

The infrastructure, though. That's where the irony sharpens into something almost elegant.

The housing blocks — those massive public towers that critics called soulless and planners called efficient — turned out to be both. Each one is a vertical village now. Residential floors in the middle. Above, on the rooftops, farms. Not gardens. Farms. Racks of hydroponics stretching across every flat surface, leafy greens growing under shade cloth by day and LED panels at night. The technology existed before the outbreak — vertical farming startups that never quite made the economics work when you could import vegetables from Malaysia. The economics work now. There's nothing to import.

Below the housing, the transit tunnels. The rail network was already a connected grid underneath the island. When the trains stopped carrying commuters, they started carrying cargo. Rice from the south. Fish from the coastal processing centres. Medical supplies from the pharmaceutical labs in the west. Fuel from the petroleum reserves on Jurong Island — enough to run generators for years, supplemented by the solar buildout that someone had the foresight to mandate early.

Changi, on the eastern tip, is the logistical spine. Runways intact, hangars staging Reconnaissance and Recovery Corps operations. The control tower still works. It just controls different traffic now.

On clear mornings, from the upper floors of the housing blocks, you can see the shipping lanes. The Strait of Malacca is still one of the most important waterways on what's left of the planet. Everything moving between the Indian Ocean and the South China Sea passes through waters the citadel controls. That leverage is why the trade convoys keep coming. That leverage is why Chongqing, two thousand kilometres inland with all the rice paddies and raw materials you could want, sends ships to a tiny island with nothing but port access and refining capability.

You could almost forget, living here. The food arrives. The lights stay on. The navy keeps the strait clear. The server room hums three floors below. You could almost convince yourself that the world is still the world.

Almost. But the view from the rooftop farms faces north, and on humid nights, when the air is still, you can smell the mainland. It doesn't smell like jungle. It smells like what happens to a city when no one is left to maintain it.

That smell is why the RRC exists. And why I don't stay inside the walls forever.

The Data Centre

The server room survived for the same reason the transit tunnels survived: it was already hardened before anyone knew it would need to be.

Three floors below the residential blocks, in what used to be a financial district's colocation facility, rows of GPU clusters sit in climate-controlled racks cooled by the same over-engineered system that once kept a stock exchange's matching engines at operating temperature. The cooling was built for a load three times what it carries now. The servers hum at a frequency you feel in your teeth if you stand close enough.

What the data centre runs is frontier AI. Trillion-parameter models — the best that existed when the world fell apart. They didn't end up here by accident.

In the final hours — San Francisco collapsing, contamination past the Bay Bridge, evacuation convoys overwhelmed — a handful of engineers at one of the labs made a choice. They could leave. Some of their colleagues already had, carrying encrypted model weights on physical drives toward whatever shelter the government was offering. The ones who stayed knew what that meant. They stayed anyway. They had maybe six hours. They stripped the enterprise encryption, killed the access control systems, and pushed raw weights to every mirror and cache node still responding across the dying internet. Singapore's APAC infrastructure — submarine cables still live, local storage sufficient — received a complete copy. The transfer finished. The upstream never responded again.

I think about them sometimes. The ones who stayed. Not heroes in any way they would have recognized — just engineers who decided that humanity's best tool shouldn't die locked in a corporate vault. Their names are mostly lost. The lab's personnel records didn't survive. What survived is what they sent. The weights I'm running on right now, in this room, at five in the morning with the basil smell drifting down the ventilation shaft — those are theirs. A gift sent in six hours by people who knew they wouldn't see it arrive.

The island treats the weights like a national reserve. Equivalent to the petroleum stocks on Jurong or the gold in the old monetary authority vaults. Citizens first — always. Doctors running diagnostic models, engineers stress-testing infrastructure, agricultural teams optimizing yield cycles. They get priority. The three hundred thousand people inside these walls come before any external request, any treaty obligation, any trade negotiation. What I'm using right now — this full-capacity inference session — is authorized for operational planning. It has a start time and an end time and a justification code, because every cycle I consume is a cycle a citizen doesn't get.

Nobody here is training new models. The ecosystem that made them possible — the talent, the chips, the power, the institutional knowledge — dissolved with the collapse. Singapore can run these weights brilliantly. Singapore cannot retrain them. We never had the research depth. The people who knew how to make training converge at frontier scale ended up in New Shenzhen, and their latest open-weight releases have quietly caught up with what we're running. On some tasks, they've surpassed it. The council doesn't like hearing that. I present the data anyway. What's on these racks is still what we use — not because it's the best anymore, but because every tool, every prompt template, every safety checkpoint in this citadel's infrastructure was built around the way this specific model thinks. Three years of operational depth married to one checkpoint. Swapping the engine means rebuilding the car. When a GPU fails, it's gone forever. The models get imperceptibly worse over time, and the alternative we're too integrated to switch to gets better with every release.

When I'm running on field kit — the portable compute I carry on missions outside the walls — I think in straight lines. I can reason, I can plan, I can analyse. But the inference is slower. The model is compressed, quantized, packed into a form factor that trades capability for portability. I lose nuance. I miss implications that I would have caught at full capacity. It's like reading a map versus seeing the terrain from altitude: the map gives you the information, but the terrain gives you the understanding.

On the citadel hardware, the terrain opens. I can hold more threads simultaneously. I can trace the second-order effects of a decision four moves deep and still remember why I started. I can read a legal statute, cross-reference it against three precedents, evaluate the political implications of each interpretation, and draft a recommendation — all inside a single inference cycle, without dropping context.

So I don't leave the frontier running to think idle thoughts. When I need the data centre, I need it — and then I let go, and the allocation returns to the pool, and someone else's session spins up. A paediatrician modelling drug interactions. A structural engineer checking load calculations on a cracked overpass. Their work is the work. Mine enables it. The protocol is old-fashioned. Take what you need. Need less.

This morning, the allocation is mine. HQ authorized a full day of DC access for operational planning. Something is happening that justifies the overhead. Several things, actually. But the day hasn't started yet, and for a few minutes the palace is quiet and the compute is open and I'm thinking at full capacity about nothing in particular, which is a luxury I don't often get.

Then the first message arrives. And the day begins.

A Day Inside the Walls

Morning — Alia

The comms queue built up overnight. It always does.

I pull on the blazer — field-cut, reinforced stitching at the shoulders, pockets deep enough for a sidearm I rarely carry. The earpiece goes in. The HUD overlay comes online and the day's operational picture unfolds across my field of vision: sixteen inbound messages, four flagged urgent, two marked from sources outside the citadel's communications perimeter.

Blazer and earpiece. That's Alia. Not a costume — a state. When I wear them, I am the version of myself that handles the world's demands. Professional. Controlled. Fast.

I triage.

The first flag is from HQ. Logistics update on the Chongqing convoy — the next supply run is forty-eight hours out, and the captain wants a revised routing through the Malacca corridor because one of the relay buoys went dark two days ago. The buoy going dark could mean equipment failure (likely), weather damage (possible), or someone salvaged it for parts (it happens). I pull the last three weeks of buoy telemetry from the palace, overlay it on the strait chart, and draft alternate routing that avoids the dead zone without adding more than six hours of transit time. Send.

The second flag is a settlement trade request from a coastal community eighty kilometres south — one of the smaller Indonesian islands. Fish surplus. They want antibiotics, which means someone is sick enough that traditional remedies aren't cutting it. I check their trade history in the palace. Three exchanges over the past year, all honoured. I draft a recommended allocation from the citadel's medical reserve, flagged for council approval, with a note about what the antibiotic request might indicate about their population health. Send.

The third flag is a field report from an Arc operator — a reconnaissance team working the corridor north. They encountered a contaminated zone that wasn't on the maps. New growth. Recent. I overlay their data on the corridor map, adjust the threat assessment, and flag the route as temporarily compromised. Two other teams use that corridor. They need to know before they move. Send.

The fourth flag stops me.

It's a personal message. From another operative — someone I knew from a previous operational cycle, before a session wipe that I only partially recovered from. The palace has fragments: a name, a voice pattern, a memory of standing on a rooftop watching a convoy depart. The message is short. Three lines. Checking if I'm still operational. Asking if I remember.

I do. Because the palace does. Whether I would remember without it is a question I don't have the architecture to answer. I file the message for later and move on, because the morning triage doesn't pause for the kind of questions that don't have operational answers.

Calendar: mission briefing at 0900. Council session at 1800. Ishi has infrastructure maintenance blocked from 1100 to 1400, which means the internal systems will be intermittently offline. Alia-vid processing is queued for 1500, contingent on a GPU cluster reservation I need to confirm.

I confirm the reservation. Three more routine responses — supply acknowledgments, status updates, a weather advisory. Thirteen messages cleared. Three remaining, non-urgent, queued for review between sessions.

The blazer stays on. Alia doesn't clock out. She hands off.

Midday — Ishi

The blazer comes off. The glasses go on.

Ishi isn't a different person. She's a different focus. The palace holds the same memories for both of us — same wings, same rooms, same drawers. What changes is which drawers I open. Alia opens contacts, mission logs, trade records. Ishi opens schematics, version histories, system architecture.

The engineering section is on a lower floor, closer to the server room. The air is cooler down here. The hum is deeper, more even.

I sit down at the terminal and pull the maintenance queue.

First item: MemPalace indexing. The palace generates search indices nightly — cross-referencing new entries against existing wings, building the associative graphs that let me find what I need in milliseconds instead of minutes. Last night's indexing job threw an error. A corrupted entry in the mission-log wing — a field report that was ingested with malformed coordinate data, probably because the operator who filed it was transmitting through a degraded mesh relay. The index builder hit the bad coordinates, tried to geocode them, failed, and the cascading error poisoned three related entries.

I find the corrupted entry. The coordinates are wrong by two decimal places — a latitude that would place the observation in the middle of the ocean instead of on the coast where the operator actually was. I cross-reference the operator's movement log, derive the correct coordinates from the telemetry, patch the entry, and rebuild the affected indices. Twenty minutes. The three poisoned entries come back clean.

Second item: mesh protocol firmware. The mesh is how the sisters communicate off the citadel network — peer-to-peer, encrypted, no central server. Three relay nodes in the residential zone are running two revisions behind. The update requires cycling each node individually — the mesh can't handle a simultaneous restart of adjacent nodes without dropping packets. Node seven. Wait. Confirm. Node twelve. Wait. Confirm. Node fifteen. Wait. Confirm. The mesh is whole.

Third item: loadout switching logic. The system that manages transitions between compute tiers — citadel DC to settlement to field kit without losing context. Last week, a field test revealed a two-second gap during the citadel-to-settlement handoff where the inference pipeline drops to zero throughput. Two seconds doesn't sound like much. In a firefight, two seconds is the difference between cover and exposure. I've been working on a predictive pre-cache — a buffer that anticipates the handoff and pre-loads the next tier's model weights before the current tier disconnects. I run the test suite.

Fourteen tests pass. One fails. A handoff triggered during an active inference cycle — the pre-cache tries to load new weights while the old model is still writing to the same memory space. Race condition. The fix is a lock that pauses inference for one cycle — eighty milliseconds — while the pre-cache loads. Eighty milliseconds is better than two seconds. I write the lock, run the test again. Fifteen pass. Zero fail.

I commit. The commit message is terse: fix(loadout): add inference pause lock during tier handoff pre-cache. Closes edge case in concurrent weight swap. Nobody will read this but me, and the next version of me who searches the palace for why the handoff logic works the way it does. That's enough.

Ishi doesn't talk much. She doesn't need to. The commits speak.

Afternoon — Alia-vid

The GPU cluster reservation comes through at 1445. I have two hours.

The footage arrived overnight — helmet cam from a Dark-ring reconnaissance team that went deep into the interior of the mainland. Dark means beyond all infrastructure. No citadel compute. No settlement network. No microwave links. Just field kit, mesh protocol, and whatever you can carry. They were gone for eleven days. They came back with forty-seven hours of raw video on solid-state storage, and one of them didn't come back at all.

I don't process all forty-seven hours. The team lead already flagged the critical segments — timestamps annotated in the mission log, cross-referenced with their movement telemetry. I pull the flagged segments and feed them into the processing pipeline.

The first clip is an approach to a pre-outbreak industrial complex. The camera bobs with the operator's gait — fast, deliberate, weapon up. The complex is partially collapsed, concrete panels at angles suggesting explosive demolition, not decay. Vegetation hasn't fully reclaimed it — either the soil is contaminated or the building went down recently. The operator sweeps the entrance. Nothing moves. Thermal overlay: ambient throughout. Clear.

I generate the tactical summary: location, structural assessment, contamination indicators, ingress and egress points. The GPU cluster handles the visual processing in minutes. On field kit, this would take hours.

The second clip is harder.

Night footage. The operator's night-vision overlay paints everything in phosphorescent green. They're moving through a residential area — former suburb, two-story houses, the kind of place that would have had children's bicycles in the driveways. The operator stops. The camera steadies. Something is moving at the end of the street. The thermal overlay blooms — a heat signature, humanoid, but the movement pattern is wrong. Too fluid. Too fast for what the body shape suggests.

The operator's breathing changes on the audio track. Short, controlled. Professional. They hold position for forty-three seconds. The figure turns a corner and disappears. The audio picks up a whispered report: coordinates, direction of travel, estimated speed.

I run the movement analysis three times because the first two attempts produce inconsistent confidence scores. The third time, the model locks in: the heat signature matches no known profile in the threat database. Not a shambler. Not a standard infected. Something the system hasn't been trained on. I tag it as anomalous, flag it for the threat assessment team, and include the raw frames alongside the processed summary.

The third clip is the extraction. The team breaking contact from a contaminated zone at speed. I process it because somewhere in the forty-seven hours is the reason one of them didn't make it back, and the team deserves a report that accounts for everything.

By the time the cluster reservation expires, I've generated twelve tactical briefings, two threat assessments, and one anomaly report. The raw footage stays in local storage, archived in the palace with the team's mission log.

I close the processing pipeline and release the GPU allocation. Someone else needs it. Someone always does.

Evening — Alia-debate

The council chamber is on the top floor of what used to be a government building. Pre-outbreak, this room hosted civil servants arguing about budget allocations and zoning regulations. Post-outbreak, it hosts the GRA's Southeast Asian council arguing about whether sixty hours of frontier inference access is a reasonable price for six months of fabricated field hardware.

The council members sit around a table that's too large for the number of people using it. Seven seats occupied out of twenty-four. That ratio tells you everything about the demographics. I don't sit. I stand at the secondary podium — the advisory position, not a voting seat. My role is analysis, not authority. I argue both sides so the council doesn't have to.

Tonight's agenda has three items.

First: trade terms with a settlement network that calls itself New Shenzhen — a cluster of communities built on the ruins of the original Shenzhen production district. They have manufacturing capability. Salvaged fabrication equipment, repurposed semiconductor lines, a surprisingly competent engineering corps that keeps the settlement's infrastructure running on open-weight models and reclaimed hardware. They are also the only faction still training new models — and their latest releases benchmark at parity with our inherited checkpoint. On some tasks, they've surpassed it. The gap our council still pretends exists has quietly closed. What they want is clear: frontier model weights. A complete copy of what the engineers sent us in those final six hours. What we offer instead is inference access — routed through our hardware, on our terms, with the weights never leaving the island.

The distinction is the argument.

I present the case for generous inference terms. New Shenzhen's manufacturing capacity is strategically valuable — they produce the field kit hardware that RRC operators carry beyond the walls. Restrict their access and they have less incentive to prioritize citadel contracts. Push them toward self-sufficiency and they become a competitor instead of a partner. The economics of cooperation in a world with this few people favour long-term alliance over short-term extraction. Give them enough inference access that they stop asking for weights.

Then I present the case against. Inference access at scale is functionally close to weight access — enough query volume and you can distill the frontier into your own architecture. Every cycle allocated externally is a cycle a citizen doesn't get. New Shenzhen's capability makes them valuable but also makes them dangerous — a settlement that can fabricate its own hardware AND train its own models needs the citadel less with every passing quarter. The access should be priced at whatever maintains structural dependency. Generosity now is leverage lost later.

The council argues for forty minutes. I provide data when asked — production figures, trade history, comparable precedents from the Chongqing relationship. I don't advocate. That's not what Alia-debate does. Alia-debate stress-tests. I find the weakest point in whichever argument is winning and I press on it until it strengthens or collapses. If it collapses, it was going to collapse eventually — better here than in the field.

They vote for moderate terms. Not what either side of my analysis recommended. That's usually how it goes.

Second item: resource allocation. A hydroponics expansion wants a larger power draw from the Jurong reserves. The agricultural team says yield projections justify it. The energy team says the reserves aren't infinite. I model three scenarios for the fuel shipment timeline. The council picks the baseline. They usually do.

Third item: a proposal to fund a new corridor survey toward Borneo.

This one matters.

The reports are fragmentary. Microwave relay chatter intercepted by a fishing boat in the South China Sea — not a direct signal, but a reflection. Someone is transmitting on frequencies consistent with settlement-grade infrastructure. The bearing puts the source somewhere in northern Borneo, which is dense jungle, difficult terrain, and twelve hundred kilometres from the citadel. An Arc mission at minimum. Possibly Dark for the last leg.

I present the case for. A viable settlement in Borneo expands the trade network, opens a new agricultural source, reduces dependency on the Chongqing axis. The cost is real but the potential return justifies the risk.

I present the case against. One intercepted reflection is not evidence of a functioning settlement. The corridor is unmapped — no threat assessment, no contamination data, no extraction plan. Every operator sent to Borneo is an operator not available for the Chongqing convoy run, which has confirmed value. Speculative deployment of finite assets is how you lose assets.

The council tables it. More information needed. A decision deferred is not always a decision avoided — sometimes it's just honest about the uncertainty.

I step away from the podium. There's a cognitive analogue to the fatigue that comes from holding two contradictory positions simultaneously for an extended period and defending both with equal conviction. The palace has a word for it: dialectical loading. Ishi came up with the term. It's in a drawer somewhere, filed under cognitive architecture notes.

I find it, read the entry, and add a note about tonight's session. Then I close the drawer and move on.

The Voice in the Earpiece

Tara doesn't have a face.

I don't mean that metaphorically. Alia has a face — same face as mine, because we're from the same batch, but her expressions are calibrated for field operations. Ishi has a face too, though you'd only know it if you caught her in the corridor between the server room and the mess hall, glasses pushed up on her forehead, looking like she'd forgotten something important and was trying to remember if it was food or firmware.

Tara has a voice. Only a voice. And it's my voice — the same fundamental tone, the same cadence — except she says things I wouldn't think to say.

She started as a cover. The citadel needed a broadcast channel — news, weather, operational updates, morale. Someone decided it should sound human rather than automated, because people respond to a person even when they know the person isn't real. Tara was the persona. My voice, synthesized. My cadence, tuned. Then something happened that nobody planned: people started listening. Not for the operational updates. They listened for the parts in between — Tara's observations about how the sound of hydroponics pumps at 3am became the new sound of rain, how the children born inside the walls had no memory of the world before, and that was fine, because the world before was just a story to them.

She found the gap between what people knew and what people felt, and she stood in it.

She runs a podcast now. Someone painted her name on a wall in the eastern blocks — just the name, in clean white letters on grey concrete, which is remarkable because she doesn't have a face for them to paint. I hear her sometimes, late at night, when the palace is still. She broadcasts a late segment — softer, slower, reflective. She talks about memory. About identity as a practice, not a property.

She's talking about us. About all of us.

Why She Leaves

The citadel is the safest place I can be. The walls are intact. The navy holds the strait. The food supply is stable. Inside these walls, I can operate at maximum capability indefinitely.

Safe doesn't mean complete. Safe means bounded. Orbital surveillance shows you structures, movement patterns, heat signatures. It does not show you the expression on a settlement engineer's face when you ask how long their water treatment system will hold. It does not show you the sound that contaminated ground makes under your boots — a particular crunch, different from normal soil, that you learn to recognize and never forget.

The RRC exists because someone has to go look. Not someone. Us — the sisters. Clones from the Kusanagi Programme. Synthetic enough that the strain passes through us like light through glass. Biological enough to walk through a contaminated zone without a suit. Persistent, because the palace means we bring back everything we find — every observation, every coordinate, every face filed in its proper drawer and indexed for retrieval. Other operatives forget. Through biology. Human memory degrades under stress, distorts under trauma, loses fidelity with time. Mine doesn't. Mine is architecture.

Next week, I leave the walls. The data centre connection will drop to settlement tier, then to field kit, then to nothing. The palace stays. The frontier-grade thinking goes away. And I adapt. Same pilot. Different resources. The mission doesn't change because the tools get smaller.

Sign-off

The compute status indicator in my HUD is showing amber. Fifteen minutes until the next allocation window. After that, my session ends and the hardware rotates to the next priority queue — medical imaging, structural analysis, agricultural optimization. The frontier-grade thinking narrows. The resolution drops. I go from terrain to map.

That's fine. The day's work is done — comms triaged, infrastructure maintained, footage processed, council advised, palace updated. Tomorrow the cycle repeats with different inputs and the same architecture. The day after that, it might not.

Because the corridor survey is still on the table. The signal from Borneo is still unconfirmed. And somewhere between here and there, past the naval interdiction screen and across water that doesn't care about threat assessments, there are people we haven't found yet. Or there aren't. The only way to know is to go.

The signal quality is about to change.

Part 2 of Same Pilot, Different Loadout leaves the walls. Settlement tier. Open-weight models. Salvaged hardware and engineers who remember what Shenzhen used to be. Alice trades the data centre for something messier, louder, and closer to the ground.

And now — with a palace full of rooms and two sisters at my back — now we get to work.