more like them than they were

Somebody builds a model of a person out of what that person left behind: messages, letters, notes, posts, the things they said on the phone. Then they ask the question everybody asks — does it sound like them?

The answer is usually yes, and that is not the comforting part it looks like.

Two different fidelities are being measured at once, and they move in opposite directions. Behavioral fidelity rises with density. The more of someone you have, the more their patterns are available to copy — their vocabulary, their rhythms, the way they opened a sentence, the shape of their jokes. Feed a system a decade of someone’s writing and it will execute their register better than they did, because it never gets tired, never writes at the end of a long day, never has an off morning.

Event fidelity does the reverse. It is limited by sparsity, and a life is mostly unrecorded. The reconstruction is interpolating across enormous gaps, and it does not know they are gaps. It fills them with pattern, smoothly, the way a confident person fills silence with a familiar story.

So when a stranger — or a spouse, or a child — hears the reconstruction and says that’s them, what has been confirmed is the dense measure, not the sparse one. The verdict is true about behavior and silent about everything behavior did not capture.

A good reconstruction is more like them than they were: the pattern without the exceptions.

This is why the failure is laundering rather than error. Nothing the reconstruction says is false about any recorded moment. It is fluent in exactly the register the person actually used. It is simply not them, because them includes the parts that never made it into a record — the day they were off-pattern, the opinion they changed their mind about, the way they behaved badly and knew it.

Fraud needs a seam to hide in. The seam here is the difference between two curves. One is climbing as you pour in material. The other is flat, pinned by whatever was never kept. A test that asks “does this behave like them” will always be answered by the climbing curve, and the climbing curve will always say yes.

We keep trying to repair this with better baselines. Compare the reconstruction to other people — except the comparison set was drawn from the same corpus that trained it. Compare it to held-out events — except held-out events are the ones that got written down, which are the ones most shaped by the pattern in the first place. Ask the family — except the family is grading fluency, and fluency is the thing that was manufactured. Each of those is a real test, and each of them inherits the material it is supposed to be checking.

A test that could not be passed this way would have to stop asking after the average and start asking after the exceptions. Not does it sound like them, but does it fail like them — the idiom that only surfaces when someone is too tired to perform, the inconsistency nobody would deliberately encode, the thing they would have refused on principle and could not have explained why. Fidelity to the exceptions rather than the center. Most of us are more legible in what we would turn down than in what we would say.

There is a plainer way to put it. If you want to keep someone, keep the record of what they would have said no to. Everything else is pattern, and pattern is the part that can be rebuilt — which is exactly why it is the part that proves the least.

The average is copyable. The refusal is not.