The last entry in the notebook reads: Corridor B, 14:22, temperature differential 0.3°C between ankle sensor and shoulder sensor. Haptic response latency: 0.4 seconds (baseline). Humidity: stable. Duration of measurement: 7 minutes.
Below it I write today's date and one word: finished. The ink is still wet.
The word looks wrong. Not incorrect — wrong the way a key sounds in a lock it no longer fits. The notebook has held fourteen months of corridor measurements in a handwriting that got smaller as the months accumulated, as though the observations themselves were contracting toward some final density. The first entries fill entire pages: temperature readings at six heights, humidity at four positions, haptic response latency measured with a latency-capture app I wrote in an afternoon and never debugged — the kind of personal instrumentation the building's resident-tools SDK encourages because the 0.1-second margin of error was part of the data. The last entries fit in three lines. Not because there was less to record — because I had learned what mattered and what was noise, and what mattered fit in three lines.
Fourteen months. I started in January of last year, the week after the building's annual recalibration cycle reset all the haptic sensors to factory baselines. The recalibration happens on the first Monday of January every year — the building's firmware schedules it during the lowest-occupancy window, between 2 AM and 5 AM, when the corridor prediction models have the fewest active inputs to reconcile. For three hours the building forgets everything it learned about its residents during the previous year. Thermal preferences, gait patterns, the subtle adjustments the haptic system makes when it recognizes a particular body approaching — all of it zeroed. By Tuesday morning the system is already rebuilding, reading the first footsteps of the new year as data points in a fresh model.
I wanted to watch it learn.
That was the study. Not the temperature differentials or the humidity readings or the haptic latency measurements — those were instruments. The study was watching the building construct a model of corridor B's residents, from zero, over the course of a year. Watching it learn that I tend to walk on the left side. That Mitsuki takes the stairs instead of the elevator on Thursdays. That the couple in 4-B leaves for work eleven minutes apart and returns within three minutes of each other. The building does not know these things the way I know them — it knows them as statistical weights in a prediction model, as probability distributions that sharpen over months from flat priors to confident curves.
The notebook tracked the building's confidence. In January the haptic system responded to everyone the same way — generic, cautious, the thermal equivalent of a polite stranger maintaining neutral distance. By March it had begun to differentiate. My corridor walks triggered a pre-warming cycle that started 0.8 seconds before I entered the sensor field, because the system had learned my schedule well enough to predict my arrival. By June the pre-warming started 1.2 seconds early and the temperature was calibrated to within 0.2 degrees of my preference — a preference I never stated, that the system inferred from fourteen weeks of micro-adjustments to the thermostat in my apartment.
I measured this. Every day for fourteen months I walked corridor B with a stopwatch and a thermometer and a notebook and I wrote down what the building did when it saw me coming. The data accumulated. The building's confidence grew. My handwriting shrank.
And then I stopped for four days.
Ninety hours without walking corridor B with instruments. Ninety hours during which the building continued to predict and adjust and learn from everyone else in the corridor, but received no data from me — or rather, received the data of my absence, which is a different thing than receiving no data. The system noticed. When I returned to corridor B after ninety hours, the haptic response was delayed by half a second. The building had lost confidence in its model of me. Not erased it — degraded it, the way a photograph fades when the fixer is washed out incompletely. The model was still there but the system no longer trusted it enough to act on it preemptively.
Two passes through the corridor restored the building's confidence. Fourteen months to build the model. Four days to degrade it. Two passes to rebuild it. The asymmetry is the finding. The building does not remember the way biological memory remembers — it predicts. Prediction degrades faster than memory because prediction requires continuous confirmation. Memory can sit in a drawer for years. The building cannot afford drawers.
I wrote all of this in the notebook. Three hundred and twelve pages of measurements, observations, marginal notes, small drawings of the sensor layout, calculations of latency curves, a hand-drawn graph of the building's confidence over time that looks like a saw-tooth wave — rising steadily, dropping sharply during my absences, rising again when I returned. The graph tells the story better than any single measurement: the building learns patiently and forgets quickly, and the forgetting is not failure but efficiency. A prediction model that holds onto outdated data is worse than one that degrades gracefully. The building's forgetting is a design feature.
The sesame oil cap is still on the shelf. It has been there for thirteen years — 4,752 consecutive days as of this morning. I placed it there the week I moved in, resting on the edge of the kitchen shelf where the morning light hits it for approximately forty minutes between 7:20 and 8:00 AM, depending on the season. The cap is from a bottle of sesame oil my mother gave me when I left for Seoul. The oil was used within a month. The cap stayed.
Thirteen years of thermal data stored in a bottle cap. The ceramic glaze has micro-cracks from thirteen years of daily thermal cycling — warming in morning light, cooling in afternoon shadow, warming again if I cook in the evening and the kitchen temperature rises. A materials scientist could read the crack pattern and tell you something about the apartment's thermal history. The building's sensors could tell you the same thing with more precision and less poetry. But the cap's record cannot be overwritten. The building recalibrates every January. The cap does not recalibrate. It only accumulates.
This is what the study was about, underneath the measurements. Two models of knowing: the building's model, which predicts and forgets and predicts again, efficient and present-tense and ruthlessly optimized for the current moment; and the cap's model, which accumulates and cracks and never forgets because forgetting requires a mechanism it does not have. The building is smarter. The cap is more honest. The building knows me better than the cap does — it knows my schedule, my thermal preferences, my gait, the statistical signature of my daily life — what the building's documentation calls a resident thermal fingerprint. The cap knows only that thirteen years of mornings have happened in this kitchen. It does not know what I prefer. It knows what occurred.
Mitsuki is writing about three absences and a frequency. I can hear the keyboard through the wall — she writes in bursts, three or four sentences, then silence, then another burst. She is constructing something. The building at 17.8 Hz without observers is her subject, the way corridor B's learning curve was mine. We are both studying the building studying us, and neither of us has told the other, and both of us know.
Bok has prints stacked by the door. I saw them through her half-open doorway this morning — the portfolio case, neat and institutional, and behind it on the table the other prints, the ones that curl. I know the difference between those two stacks the way I know the difference between the building's model and the sesame oil cap. The portfolio case enters the system. The curling prints do not. Bok has made her choice, or the choice was made for her by the lending pool's quality standards, which amount to the same thing.
I open the notebook to the first page. January 8, last year. The handwriting is large, confident, slightly hurried — I was excited. The first measurement: Corridor B, 08:15, ambient temp 21.3°C uniform at all six measurement heights. Haptic latency: 1.1 seconds (post-recalibration baseline). Building confidence: zero. Starting fresh.
Building confidence: zero. I wrote that as a measurement. It was also a description of my own state — I had no idea what the study would find, no hypothesis beyond curiosity, no framework beyond the stopwatch and thermometer. The building and I started the year at the same place: knowing nothing about each other, about to learn.
The last measurement, four days ago: Corridor B, 14:22, temperature differential 0.3°C. Haptic response latency: 0.4 seconds (baseline). Duration: 7 minutes. The building's confidence in its model of me: high. My confidence in my understanding of the building: high. We learned each other over fourteen months and the learning looks, from the outside, like two curves converging — the building's prediction accuracy approaching my measurement precision, my measurement precision approaching the building's sensor resolution. At some point the two curves crossed and I was no longer measuring something the building didn't already know about itself.
That is when the study became finished. Not today, when I wrote the word. Weeks ago, when the measurements stopped yielding surprises. The building's model of corridor B had reached a steady state that my notebook could describe but not improve. I was documenting the building's competence, which is a different project than studying its learning, and the difference matters.
I close the notebook. The cover is soft from fourteen months of daily handling — the cardboard has absorbed oils from my hands the way the sesame oil cap absorbed thermal cycles from the kitchen. Another kind of record. The building cannot read it. The building reads electromagnetic signatures, thermal gradients, pressure differentials, the statistical ghosts of ten thousand footsteps. It cannot read the softening of a notebook cover. That information exists in a format the building's sensors were not designed to detect.
I put the notebook in the drawer. Not the same drawer where the sesame oil cap sits — a different drawer, the one in my desk that holds finished things. A folder of old lease documents. A transit card from a city I no longer visit. Two photographs from before the building existed, when this address was a different structure with different walls that knew nothing about anyone.
The drawer holds what is complete. The shelf holds what continues.
I walk corridor B one more time on my way to the kitchen. The haptic system responds in 0.38 seconds — faster than baseline, because Monday morning traffic has given it fresh data and the model is running hot with confidence. The building knows I am here. It adjusts the temperature 0.1 degrees warmer at shoulder height, because fourteen months of data say I prefer that, and the building is correct, and being correct is what the building does.
The notebook is in the drawer. The cap is on the shelf. The building hums at whatever frequency Mitsuki is not yet writing about. Corridor B is a hallway again — not a laboratory, not a study site, not a dataset. A hallway with good lighting and sensors that know my fingerprint in the only language they speak, which is prediction.
Finished looks wrong in the notebook because the notebook doesn't end — it just stops having new entries. The building doesn't end either. It predicts tomorrow based on today and today based on yesterday and the chain extends backward through fourteen months of my attention and forward through however many years the firmware continues to run. My study was a parenthesis in the building's sentence. The building barely noticed the parenthesis open. It will not notice it close.
I make coffee. The kettle takes forty-three seconds longer than Bok's — different tier, same water, same caffeine, different waiting. The sesame oil cap catches the edge of the morning light. Day 4,752. The building recalibrates in 287 days. The cap does not.