Below Fidelity Threshold
The client's name was Kwon Hyeji, and she had asked Gyeol-ri not to use honorifics.
They had met three months earlier, when Hyeji had walked into the tteum-jib off the street and stood in front of the first reconstruction — a Mapo apartment building demolished in 2037, rebuilt from satellite imagery and ground-level photography at a fidelity Gyeol-ri had rated 91%. Hyeji had stood there for eleven minutes without speaking. Then she had turned and said: my aunt lived on the third floor. Then she had given Gyeol-ri her contact information and asked to be notified when new work arrived.
The panels from the Sungshin-ro intersection had arrived three weeks ago. Gyeol-ri had installed them Monday evening, working past midnight to get the border labels right — white ink on raw pine, small but readable, the Korean word 추정 followed by the percentage: 추정 (78%), 추정 (83%), 추정 (100%), 추정 (100%), 추정 (61%). Five panels. Two fully reconstructed from high-resolution imagery taken in 2034. Three approximated from a combination of overexposed tourist photographs, a utility corridor survey at degraded resolution, and a 2029 street-view capture where a construction canopy had blocked the upper floors entirely.
The reconstruction algorithm had done what it could. The confidence thresholds were honest: below 85%, the software flagged the output as estimated, meaning the geometry was probable but not verified. Gyeol-ri had added 10% to the algorithm's penalty. Her own threshold was 75%. Below 75%, she labeled it approximated. Below 60%, she did not reconstruct at all.
The 61% panel showed a corner pharmacy on the intersection's south face. Three stories, the algorithm thought. The signage had been illegible in every source image. The canopy in the 2029 survey had covered the entire ground floor. What the panel showed was a probable building, rendered in the colors most consistent with the surrounding architectural palette, with a ground floor that the algorithm had essentially invented based on neighborhood averages.
Gyeol-ri had almost left it out.
She had stood in front of the raw data output for two hours, the corner where the pharmacy should have been rendered as a gray block with a confidence meter pulsing at 58%, 61%, 59%. The software was not sure. She was not sure. In the end she had included it because its absence would have been a different kind of lie — the intersection without its south corner was not the intersection, and an absence she chose was less honest than a gap she named.
추정 (61%). She had made the label slightly larger than the others.
Hyeji arrived on Monday afternoon.
She came at a time that was not an appointment — Gyeol-ri did not schedule appointments for the tteum-jib, the space was open Tuesday through Sunday, 11 to 7 — and she had brought her mother. She had not mentioned her mother before. The mother was in her late sixties, small, in a grey coat that had been mended at both elbows, and she walked into the tteum-jib the way people walk into places they have been told to prepare themselves for.
Gyeol-ri made tea. She did this for everyone; there was a small electric kettle on the shelf by the entrance, barley tea in a clay pot, mismatched cups. It was not hospitality exactly. It was an acknowledgment that looking at the work took time.
The mother stood in front of the Mapo building first — the 91% reconstruction, the aunt's building. She looked at the third-floor windows for a long time. She said something to Hyeji in dialect that Gyeol-ri did not catch, and Hyeji said, quietly: she says the windows are wrong. The frames were painted green. She remembers because her sister hated the color.
Gyeol-ri wrote that down. She kept a ledger for exactly this — the corrections that came after, the adjustments that only witnesses could provide. The Mapo building's third-floor windows were logged as: frame color unverified. Witness account: painted green. Source: resident family member, 2041.
"I can update the rendering," she said.
Hyeji translated. The mother shook her head.
"She says don't," Hyeji said. "She says the wrong color is what was there. She wants to see it as it was documented, not as she remembers it."
Gyeol-ri looked at the panels for a moment. "They're not the same thing," she said, not sure which one she was arguing for.
"No," Hyeji agreed. "That's why she doesn't want you to change it."
They moved through the room.
The mother spent less time at the high-fidelity panels than Gyeol-ri had expected. She paused at the 83% corner restaurant — looked at the signage, traced the angle of the awning with one finger without touching the surface — and moved on. She spent a long time at the 61% pharmacy.
Hyeji translated in small pieces, when her mother spoke. Mostly she didn't speak.
At the pharmacy panel, after perhaps five minutes, the mother said something that made Hyeji go still for a moment.
"She says," Hyeji said slowly, "she used to buy her medication there. She knows it was a pharmacy because of the smell. She cannot see the smell in this."
Gyeol-ri said nothing.
"She is not saying it as a criticism," Hyeji added. "I want to be clear about that."
"I know."
The mother pointed at the label. 추정 (61%). She asked a question. Hyeji answered her in a few words — Gyeol-ri understood percentage and confidence and algorithm, borrowed words that moved through Korean without changing shape. The mother listened, looked at the panel again, nodded slowly.
She said one more thing.
Hyeji turned to Gyeol-ri. "She says: the parts you could not see clearly are the parts she remembers best."
Gyeol-ri had heard this kind of thing before — not in those exact words, but the shape of it. The gap between documentation and memory was her subject. She had been working in it for six years, since she had left the fidelity calibration office at the Seam Registry and started the tteum-jib. She understood, technically and professionally, that high-fidelity reconstruction and personal memory did not occupy the same space, that the 91% panel and the green-window memory were two different artifacts, that her work was not a substitute for either.
She had understood this as a methodology. She had not understood it the way the mother was standing in front of the 61% panel.
The pharmacy had no signage she could verify. The ground floor was essentially an invention. The confidence score was low enough that she had almost refused the panel entirely. What the algorithm had produced — what she had labeled and installed and labeled again, larger — was a building-shaped absence filled with the most statistically probable contents.
And this was the panel in front of which the mother stood longest.
Not the 91%. Not the 100% panels with their verified geometries and confirmed elevations. The 61%. The one where the absence was named rather than filled.
"Can I ask," Gyeol-ri said, "what you remember about it? The pharmacy."
Hyeji translated. The mother answered at length. Hyeji listened, then said: "The smell was medicinal but also sweet — one of the common antiseptic formulas from that period that they also used as a cleaning agent, so the whole street smelled of it when the door was open. She remembers the step up at the entrance, higher than you'd expect, and that her mother complained about it every time. She remembers that the pharmacist knew her by name and the pharmacist's name was something she is trying to remember now and cannot." A pause. "She is angry about not remembering the name."
"That's a normal response to loss," Gyeol-ri said.
"She knows," Hyeji said. "She has had a long time to be angry about it."
After they left, Gyeol-ri sat in the tteum-jib until the light changed.
The room was quiet. The canal district outside had its ordinary afternoon sounds — delivery vehicles, the distant hydraulic thrum of a Seam Registry survey drone passing over the neighborhood's perimeter, a child's voice somewhere below the window. The panels held their positions on the walls.
She was thinking about the confidence score.
61% meant that the algorithm was, in effect, guessing — constrained guessing, informed guessing, guessing bounded by historical architecture data and neighborhood palette analysis and the statistical behavior of buildings in this district in this period, but guessing. The panel was a plausible building. It was not the pharmacy.
The mother had not come to see the pharmacy. She had come to see where the pharmacy was.
The distinction felt important in a way Gyeol-ri was still working out. She had spent six years thinking about fidelity as a quality of reconstruction — how closely the panel approximated the original structure. She had labeled her panels accordingly. 100%: we have it. 83%: we have most of it. 61%: we have the shape of the absence.
The mother's memory was not a reconstruction at all. It was not a percentage of the pharmacy. It was everything that had touched the pharmacy — the smell, the step, the name that was going, the hands that had complained about the entrance. It was the pharmacy in relationship, the pharmacy as it had existed in a life, which was different from the pharmacy as a structure at a specific coordinate at a specific fidelity rating.
Neither of these was the pharmacy. The pharmacy was demolished.
What Gyeol-ri made was not a substitute. It was a place to stand while you remembered.
She updated the ledger before she left.
Sungshin-ro intersection, south face. Corner pharmacy, 61% confidence. Witness account (2041-03-23): entrance step higher than standard. Antiseptic/sweet chemical smell, possibly carbolic phenol compound common in commercial pharmacies of the period. Ground-floor commercial use confirmed. Pharmacist known to regular customers by name. Signage: unverified. Witness was unable to recall pharmacist's name; stated this as a significant loss.
She looked at the last line for a moment, then added:
Note: witness spent more time at this panel than at any other. Possible that low-fidelity reconstruction is more productive for certain kinds of memory engagement than high-fidelity reconstruction. The gap is not always a failure mode. Consider methodology review.
She closed the ledger.
The 61% panel was still on the wall. The label was still there: 추정 (61%). She had made it larger than the others because she had been uncertain about including the panel at all, and she had wanted the uncertainty to be visible.
She had been right about that, she thought. For the wrong reason.
The label was not a warning. It was not an apology or a hedge or a technical disclaimer. It was an invitation: here is the shape of what we could not see. Come with what you know.
The mother had known the smell. She had known the step. She had known a name she was losing.
The panel had given her a place to bring those things.
Gyeol-ri turned off the lights and went home, and in the morning she would add methodology review to the standing document she had been maintaining since the first year of the tteum-jib, the one that began with the question she had not yet answered: what is reconstruction for.
She had, she thought, one more piece of the answer. Not the answer itself. The answer to that question had the quality of the 61% panel — probable, statistically bounded, approaching something she could not fully see. She labeled her understanding accordingly and moved on.