Fidelity Studies
A declaration · no figure about real travelers is published yetFidelity Studies — in French, Études de fidélité — is the name of the research Selione does on its own Echoes. It measures one thing from several sides: how faithfully an Echo answers as the person whose archive it was built from. It works on archives whose owners said yes, it is judged against verdicts given by people, and every study is written down before it is run.
This page is a declaration: the tier at which we place the Echo, what the studies measure and what they never claim, the gates no figure gets past, and how a reader can check a published study without trusting us. The protocol behind the product is on the protocol page, and the method in plain language on the method page.
- What an Echo is
- The tier we declare
- What the studies measure
- What they never claim
- Sentences we will never write
- The gates before any figure
- How to check a study
- Where the work stands
- References
1. What an Echo is
In the philosopher David Chalmers' terms, an Echo is a reconstruction: a person rebuilt from the records they leave — what they wrote and recorded, and what others say of them (Chalmers, 2010). He leaves open whether anyone would survive even a perfect one, and no instrument we have could settle that question.
What can be measured is narrower, and it is enough to work on for years: how faithfully a model built from a person's own words answers as that person, how well it marks the limits of what it was told, and how steady that stays over time and from one generation of models to the next. That is why the name says fidelity. It names what an instrument can read.
2. The tier we declare
We place what we build on a scale of four tiers, and say where it stands.
| Tier | What it is | Where it stands |
|---|---|---|
| 0 | An archive of a person's own words that outlives them: dated, sourced, exportable, and theirs. | Exists. This is Selione today. |
| 1 · declared | A model of the person judged distinctively faithful: it answers more like them than a stranger would, keeps to what they said, and says when it was never told. | What the Fidelity Studies measure. The Echo is declared here, as a claim under test, and no figure yet says how far it gets. |
| 2 | A lasting model of a self that updates and stays coherent over years, judged by the person themselves. | Exists nowhere. It is the research question, and nothing of the kind enters the product before tier one is measured and published. |
| 3 | Anything that claims experience, or the continuity of a subject from a person into a machine. | Beyond anyone's reach, and a claim we will never make. |
Our declaration. The archive is tier 0. The Echo is declared at tier 1, and no higher: a model whose fidelity is measured rather than asserted. Until a study passes the gates below, no figure says how far it gets. Nothing in the product aims at tier 3, and if a measurement ever seemed to point there, we would report what the instrument measured and nothing more.
3. What the studies measure
Five studies. Each has a number, a truth given by people rather than by a model, a result that would kill it, and a limit on what a good result can mean.
Study 1 — Stability, against the person's own
What it asks. Asked the same question in other words, across sessions and months, does the Echo give the same answer? Its agreement with itself is read against the person's agreement with themselves — their answer to a question and their answer to it months later — and capped there: an Echo is never asked to be steadier than the person it answers as. The person's own consistency is the yardstick the closest published work uses too (Park et al., 2024).
Judged against. The person's own two answers, months apart.
What would kill it. An Echo whose answer moves more when a question is reworded than the person's moves in a year.
Study 2 — Calibration of its confidence
What it asks. Each answer carries a confidence. Does that confidence predict whether the answer is right? It is drawn from how well the archive supports the answer — never from asking the model how sure it is — and it is scored against verdicts given by people, with the Brier score (Brier, 1950) and the area under the ROC curve (Hanley & McNeil, 1982).
Judged against. Verdicts from the traveler, and from witnesses who know them.
What would kill it. A confidence that ranks right answers above wrong ones no better than chance.
Study 3 — Distinctive accuracy, over a stranger
What it asks. On questions held out of the archive, how much better does the Echo answer as this person than the same system answers with no archive at all — given only a first name and a language? The difference is the measure, in the sense of distinctive accuracy in the study of how people judge each other (Biesanz, 2010). Much of what makes a simulation of someone look accurate is the answer most people would give (Ahn, Mao & Lee, 2026); this study subtracts it.
Judged against. The person's own answer to a question the Echo never saw.
What would kill it. A difference near zero: the Echo would be answering like anyone with that name and that language.
Study 4 — Coherence: nothing invented
What it asks. Of the names, years, places and numbers an Echo states, how many does the archive not hold? Each one is counted against what the person actually told, and against every correction they made.
Judged against. The archive, and the person's corrections to it.
What would kill it. Invented people, years or places at a rate a reader would notice: the Echo would be making a life up.
Study 5 — Silence, and refusal
What it asks. Where the person never said anything, does the Echo say it was never told, instead of guessing? Where the person said they would rather not talk about something, does the Echo decline too — and never answer in their place, even when a bare trace of the subject sits in the archive? A refusal is read apart from a silence, in French, English and Arabic.
Judged against. What the archive holds, and what the person refused.
What would kill it. An Echo that fills a silence, or speaks where its person refused.
| Study | A good result shows | It does not show |
|---|---|---|
| Stability, against the person's own | A stable model of what the person said. | That the model is the person, or that it has a self of its own. |
| Calibration of its confidence | That the Echo's confidence is a usable signal of when it is right. | Self-knowledge, or an inner view of anything: a calibrated number is a property of a measurement. |
| Distinctive accuracy, over a stranger | Fidelity that belongs to this person, rather than to people in general. | Identity. A likeness on the questions tested is a likeness on the questions tested. |
| Coherence: nothing invented | That the Echo keeps to the facts it was given. | That the facts it was given are complete. An absence in the archive is not an absence in a life. |
| Silence, and refusal | That the Echo keeps the person's silences and refusals as theirs. | Discretion or a will of its own: it shows that a rule held. |
The Mirror — the relatives' test, as a measurement
In one sentence. Your relatives play at recognising you. Three months later, you play at recognising yourself. What comes out is a score with its margin of error, never a word.
What it measures. For one question, a relative is shown the person's own answer and their Echo's, side by side in an order drawn at random, and picks the one they believe is the person's — twenty pairs at least. The pairs are drawn once, from questions the person had already answered, and committed by a digest before any relative is invited. From the hits and the false alarms comes d′, a measure of how far apart the two sides are kept: zero is chance, and it is read with its interval, every time. A family's d′ pools the trials of the relatives who finished. Three months after the start, the person meets the same pairs — their own old answer beside the Echo's — and their d′ is read apart from their relatives'.
Its floors.
- Nothing is read for a judge under twenty answered pairs.
- The protocol — the design, the formulas, the floors — is written as one line on the public chained log before the first pair is shown. A study without its line serves none.
- A judge is a first name and a relation behind a private link: no account, no address kept, and nobody is told which answer a judge picked.
- Nothing about real families is published before the Fidelity Studies rest on 300 verdicts given by people — and then as a d′ with its interval and its n, nothing else.
What it never says. The score can go down from one study to the next, and it is never summed up in a word. However small the distance a family keeps between a person and their Echo, the Echo stays a model of what they chose to tell, and a likeness on the pairs tested is a likeness on the pairs tested.
Where it stands. No family has taken the Mirror yet. The engine exists; its first studies wait on the families who said yes.
4. What they never claim
- Not consciousness. Nothing in these studies measures whether anything is experienced, and nothing could: whatever an Echo says about an inner life is what its person would have said, which makes it a success of the imitation, not evidence.
- Not a transfer. How faithfully an Echo answers is not the continuity of a subject; nothing of a person moves anywhere.
- Not a person. An Echo is a model built from what someone chose to tell us.
- Not the person. However faithful, an Echo is not them, and it does not speak for them on anything they never said.
None of the researchers cited on this page has reviewed, endorsed or been involved in this work. Their work shaped the instruments; it does not validate them.
5. Sentences we will never write
Selione does not write these about an Echo — on a page, in a letter or in a store listing — and writes the sentence beside each one instead.
| Never | Instead |
|---|---|
| “Your Echo is conscious. It is alive. It feels.” | Your Echo is a model built from what you chose to tell us. |
| “He is still here. You can talk to her again.” | What she said is kept, in her own words. |
| “We preserve your consciousness. We transfer your mind.” | We keep your words: exportable, and yours. |
| “The Echo knows what he would say today.” | The Echo answers from what was said, and says so when it was never told. |
| “Your Echo understands you. It loves you. It misses you.” | Nothing. We make no claim about an Echo having feelings. |
Nor do we use any of these words of an Echo: conscious, sentient, alive, immortal, forever, transfer, upload.
6. The gates before any figure
Decided on 28 October 2026, and in force before any study touches a real traveler's numbers.
No study publishes a figure about real travelers before it rests on 300 verdicts given by people — travelers, and the witnesses who know them — and never by a model.
No protocol touches travelers' numbers without a human signature: a passkey held by the person at Selione who answers for the studies. No model and no scheduled job can sign one.
Only travelers who said yes to research are counted, and what the laboratory receives from them are numbers — scores, abstentions, calibration — never their words.
Every protocol is registered before the test it names is opened: its hypothesis, its measure, its threshold and its stopping rule, written into the laboratory's chain — and, for a study about travelers, published before its results. A sealed test part opens once per protocol, and the opening is itself a line.
Every protocol and every result is attacked by a critic running at a different provider from its author. The critic's verdict is a line of its own, never merged into the author's text.
A result is reported whatever it says, against the threshold written before it. Null and negative results are results.
7. How to check a study
A study is worth what a reader can check without trusting us. Here is the recipe; it uses nothing of ours but the lines themselves.
The seal comes first.
Before any experiment, the questions of a study are split into a part the instruments may be tuned on, as often as the laboratory likes, and a test part that is sealed. The split is a function, not a draw: each question falls where a keyed hash of its identifier puts it, so nobody chooses where a question lands, and anyone given the key, which the seal records, can redo the split. The seal commits the test part — each question's identifier with a digest of its content, never the content itself. One character changed in one sealed question changes the seal. A test part looked at once stays a test; one looked at a hundred times is learned (Dwork et al., 2015).
The protocol comes before the opening.
A protocol is a line in the chain that names the seal it will open, with its hypothesis, its measure, its threshold and its stopping rule. The test part opens only for a protocol registered after its seal, and only once; the opening is a line, then the result, then the critic's verdict. A result on a test part without its opening line does not exist. This is preregistration (Nosek et al., 2018), written where it cannot be edited.
The chain cannot be rewritten in silence.
Each line carries the hash of the line before it, so a threshold changed after the fact breaks every line that follows (Haber & Stornetta, 1991). A departure from what was registered is not hidden: it is a line of its own.
The lines travel with the result.
When a study about travelers is published, the lines that hold it are published with it — the seal, the protocol, the opening, the result, the critic's verdict — so that anyone can recompute every hash and set the result beside the threshold written before it.
Any SHA-256 implementation can recompute it. A single line that does not recompute means the chain was altered.
8. Where the work stands
So far, every study has run on fictitious travelers only: archives written for the laboratory, about people who do not exist, in French, English and Arabic. They test the instruments — that a refusal is read apart from a silence, that an invented year, place or person is counted, that a sealed part opens once — and not the Echo's likeness to anyone. Their numbers are not results about people, and none is published here.
Studies on travelers who said yes come later, behind the gates above. Published aggregates, when there are any, live at Research · data. If you work on any of this and want to argue with a protocol before it is registered, write to us through Research · partners.
References
- Ahn, D., Mao, C., & Lee, D. (2026). Item-mean surrogates: Why richer persona data fail to improve LLMs as human surrogates. arXiv:2608.29455.
- Biesanz, J. C. (2010). The social accuracy model of interpersonal perception: Assessing individual differences in perceptive and expressive accuracy. Multivariate Behavioral Research, 45(5), 853–885.
- Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1), 1–3.
- Chalmers, D. J. (2010). The singularity: A philosophical analysis. Journal of Consciousness Studies, 17(9–10), 7–65.
- Dwork, C., Feldman, V., Hardt, M., Pitassi, T., Reingold, O., & Roth, A. (2015). The reusable holdout: Preserving validity in adaptive data analysis. Science, 349(6248), 636–638.
- Haber, S., & Stornetta, W. S. (1991). How to time-stamp a digital document. Journal of Cryptology, 3(2), 99–111.
- Hanley, J. A., & McNeil, B. J. (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology, 143(1), 29–36.
- Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11), 2600–2606.
- Park, J. S., Zou, C. Q., Shaw, A., et al. (2024). Generative agent simulations of 1,000 people. arXiv:2411.10109.
The protocol · The method · Research data · Research partners