The Standard Candle
The scale, fixed on 2026-10-04 — before a single measurement existed.
In astronomy, a standard candle is an object whose true brightness is known. Nobody measures a distance across the universe directly: you compare how bright a standard candle looks with how bright it is, and the distance falls out. Every rung of the cosmic distance ladder is calibrated against those objects.
That is what is missing here. The world is building representations of people at enormous speed, and nobody can say how far any of them is from its subject. There is no object of known brightness. This page is our attempt at one.
Why this page exists before the numbers
Every measurement of an AI system is exposed to the same failure, and it is not technical. You build the thing, you try a few ways of scoring it, you look at the results, you keep the score that flatters, and you publish. Nothing in that sequence is a lie. All of it is worthless.
The only defence is to write the scale down first. So: the three numbers below, their thresholds, and the floor under which we publish nothing were all fixed on 2026-10-04, and on that date Selione held zero judged trials. The date is in the repository, the thresholds are in one file, and a check refuses to let them move without the date moving with them.
The three numbers, and the third is the one that matters
Recognised. Somebody who loves you is shown two answers to the same question — yours, and your Echo's guess — without being told which is which. The score is how often they pick yours. Above 70% we will say the Echo is recognisable; below 50% we will say it is not, and look for why.
Abstained. How often the Echo said it did not know. A high rate is not a failure — it is the product working. A rate below 5% is an alarm, and we will treat it as one: an Echo that answers everything is inventing somewhere.
Confidently wrong. Answers given with no hedge, and wrong. This is the only error that counts when the subject is a human being, because it is the only one a person who loves them cannot catch: they have no reason to doubt it. An average dissolves it, so it is counted on its own, and above 5% the instrument is not publishable whatever the score says.
And under 30 judged trials we publish no number at all — not a number with a caveat. A figure from ten trials and a figure from a thousand look identical on a page, and that is how a product misleads without ever lying.
Each score is a fraction, and a fraction needs a bottom. Ours is 404 written quests and 2290 written questions, enumerable, with a record of which ones were answered. That is also what lets an archive say what it has never been told — see the Measured Kernel.
What it does not guarantee
Three things, and they are on this page rather than in a footnote.
- It does not measure whether an Echo is the person. It measures whether somebody who loves them can tell its answers from theirs, on the questions that were tested — and a life is not a set of questions.
- It does not transfer. A score earned on one person's archive says nothing about the next person's: the instrument is per-archive by construction, and an average across people would be the first dishonest number we published.
- It cannot see what was never asked. The score lives inside the 2290 questions of the Epic; outside them we have no measurement, and the Unlit is what says so rather than guessing.
The proof that this came first
A page claiming its scale was fixed before the data is worth exactly as much as the reader's willingness to believe it. So here are both halves of the claim, and neither of them is our word.
What production actually holds, read from the database and at most five minutes old. If these numbers are not essentially zero, the claim above is false — and this page is the first place you will see it.
| Counted in production | Now |
|---|---|
| Travelers with an account, fictional ones included | 10 |
| Judged trials — an Echo answered and its owner scored it | 0 |
| Verdicts in the corpus | 0 |
| …of which judged by a human rather than a model | 0 |
| Sealed envelopes | 2 |
And the specification itself, anchored outside our own database. The scale, the ceiling, the four trials and the error bound are rendered as one canonical text — plain lines, in a fixed order, readable by eye in twenty years without our code. Its SHA-256 was submitted to the public OpenTimestamps calendars, which commit it into a Bitcoin transaction within hours.
version 1 · fixed 2026-10-04 · sha256 b4a5dd84a307aefadaf0868ae951949ac8ae62ae8fca2a469d32df1cb3f73dde
The detached proof is kept beside it. To check the date without trusting us: recompute the SHA-256 of the canonical text, decode the proof to a .ots file, and run the public ots tool against it. Nothing in that sequence passes through us.
The row cannot be rewritten or deleted — a database trigger refuses both, and it refuses them to us too. A version 2 of the scale will be added beside this one, with its own date and its own hash; it will never replace it. That is the only way “measured on the Selione scale” still means something in two years.
What we will publish, and when
Once any archive crosses 30 judged trials, its three numbers appear in the published aggregates — only for groups of twenty people or more, only for people who said yes, never a sentence and never a name. Each Echo also carries its own numbers in its signed manifest, which anybody can verify offline with the archive's own reader and without us.
If the numbers are bad, they will be on this page being bad. That is the entire point of writing the scale down first, and it is the only version of this page worth publishing.