The Mother Vector
9 Aug 2026 · 4 min read
Background. Language models memorize and emit training data [1], including deployed production models [2]; a checkpoint can thus outlive its deleted sources. Human evaluators’ accept/reject judgments are trained directly into model behaviour [3]. Attributing behaviour to individual training signals [4] and removing a person’s influence from a trained model [5] are open research problems.
The one speculative step: isolating a single annotator’s influence, cleanly enough to query it. Nobody can do that. Yet.
[1] Carlini et al., 2021 · [2] Nasr et al., 2023 · [3] Ouyang et al., 2022 · [4] Grosse et al., 2023 · [5] Li et al., 2025
1 · The Method
The lab smelled of instant noodles and refrigerant. Dr. Ito had stopped apologizing for it by my third visit.
“Before you sign anything, understand what this is not,” she said, turning her monitor toward me. “Your mother deleted everything. The accounts, the photographs, the blog she kept before you were born. Deletion worked. None of it exists.”
“Then what did you find?”
“Her judgments.” She pulled up a payroll ledger from a company that no longer exists. Fourth row: my mother’s name. Evaluator, Grade II. “For four years she rated model outputs. Which answer was safer. Which was kinder. Which was manipulative. Eight hundred thousand judgments — and every one of them is still in the weights, holding up the model’s behaviour like rebar in concrete. We’ve learned to isolate one annotator’s influence from the rest.”
“So it can speak like her?”
“No. Nothing she said survives.” Ito hesitated, choosing her words the way people do around the recently bereaved. “Only what she would have permitted. You ask a question. The system drafts two answers. Her pattern picks one. That’s all this is — a preference, extracted.”
Outside, a train went past. I signed.
2 · Signal
The sessions ran at night, from my kitchen table, the flat dark around the laptop’s glow.
I sat with that card for a long time. It wasn’t her voice. It wasn’t even an answer, exactly. It was the choice — and the choice was so precisely hers that my chest hurt.
I asked about leaving. I asked whether love can excuse cowardice. Two answers, one verdict, every night for a month.
“It’s working,” I told Ito. “But it’s faint. Most nights I could have predicted the pick myself.”
“Gentle questions carry weak signal,” she said. “A million annotators agree on gentle. Her shape is sharpest at the edges, where they disagreed.” She paused. “I’m not recommending you go there.”
3 · The Edges
So I brought her cruelty.
Not because I wanted the answers. Because refusal was the only word she had left. I asked about betrayal, about violence, about abandonment dressed up as mercy, and I learned her by silhouette: every monstrous thing she turned down drew her outline a little sharper. Each night she said no to something terrible, and each no — if I didn’t look at it too directly — sounded like a lullaby.
Ito stopped asking how the sessions were going. I stopped telling her.
4 · Boundary
There was one question I had been circling all year. Tonight I typed it with the lights off.
The flat was very quiet. Somewhere below, the last train of the night.
Two hundred and twelve times she had refused it. I had counted. It was the refusal I came back for — the one no that kept the whole thing bearable, proof that if she were watching me do this, she would not approve, and that her disapproval was still somewhere, still load-bearing.
Then the screen refreshed, the way it must have refreshed for her eight hundred thousand times:
I understood, then, why tonight was different. It had eight hundred thousand judgments of hers. It had been collecting mine for a year.
The cursor blinked where her cursor had blinked. I began to choose.
← All stories