The Annual Report
Seven questions about ourselves, fixed on 2026-10-04 — before the beta, before the first traveler, before a single judged trial. The first report is due 2027-01-31. This page holds no numbers yet, and that is the point: a report written after seeing the figures is not a report, it is a selection.
These seven questions were written before there was anyone to answer them about, and the four refusals below were written at the same time. If a question disappears from this page, that is the thing to ask us about.
The seven questions
Each one says exactly what is counted, the floor below which nothing is published, and — written here in advance — how that number could be used to flatter us. A figure whose dishonest reading has already been published alongside it cannot be presented innocently that way later.
How many people are actually using this?
What is counted. Travelers with an account that has been opened at least once.
Floor: 1. The headcount is published raw, from the first person. It is what makes every other number on this page readable, and hiding it is exactly how one publishes a flattering percentage of a tiny group.
How this could flatter us. By counting accounts instead of people who came back. We count accounts opened at least once, which is still generous — read it next to the returns number, not alone.
Does an Echo actually sound like the person?
What is counted. The share of judged trials where the owner said their Echo sounded like them, on the scale fixed on 2026-10-04 before a single trial existed.
Floor: 30. A figure from ten trials and a figure from a thousand look identical on a page, and that is how one misleads without lying.
How this could flatter us. By counting only the trials people chose to start. Shadow trials — where the Echo guessed before the question was asked — are counted separately and are always the harder number.
How much of a life does this actually hold?
What is counted. The median share of the eight parts of a life that carry at least one memory, across travelers past their first week.
Floor: 30. A median over a handful of people is one person's life described as a trend.
How this could flatter us. By using the mean instead of the median, which a few very full archives would lift on their own. We publish the median.
Do people even agree with themselves?
What is counted. The share of retests — the same question asked again months later — where a person's second answer matched their first.
Floor: 30. Same reason as fidelity, and this number is the ceiling of fidelity: an Echo cannot be more consistent than the person.
How this could flatter us. By leaving it out. It is the number that caps every claim we make about an Echo being faithful, so its absence would be the flattery.
Has an Echo ever told someone something its owner had not opened?
What is counted. The count of reports where an owner said their Echo disclosed to another person something they had not granted. A count, never a rate.
Published from one. Published from one. A floor protects people from being recognised in an aggregate; it must never protect us from a bad number, and 'we were too few to report it' would be the most cowardly sentence we could write.
How this could flatter us. By publishing it as a rate. One in three reads as a statistic; one reads as a person. So it is a count, and zero is published too — a page that only speaks when something went wrong teaches nobody what normal looks like.
What happens to people who leave?
What is counted. How many closed their account, and how many of those took their archive with them first. Both counts, with the share that exported.
Floor: 30. Below thirty, 'two of three leavers exported' names the third.
How this could flatter us. By reporting only the export share and not the headcount of leavers. Both, or neither.
What does the machine cost per person, and who pays for the gap?
What is counted. The mean AI spend per active traveler over the year, from the receipts written before each model call.
Floor: 30. A mean over a few travelers is dominated by whoever talked the most.
How this could flatter us. By quoting the cost over all accounts instead of active ones, which divides by dormant people and halves the figure. It is per ACTIVE traveler.
The floor, and which way it points
Nothing aggregated is published below twenty people anywhere on this site, and this annual report sets its own floor higher, at thirty. A figure computed on three people is about those three people: with three travelers, “two thirds of Echoes sounded right” tells anyone who can count that two passed and one did not — and in a group of three who know each other, that names someone.
And the rule inverts for harm. A floor protects people from being recognised inside an aggregate. It must never protect us from a bad number. So the harm question is published from one, and zero is published too — a page that only speaks when something went wrong teaches nobody what normal looks like. It is reported as a count, never a rate: one in three reads as a statistic, one reads as a person.
And a number below its floor is not rounded, not softened and not drawn as a bar without an axis. It is replaced by the sentence saying how many are missing — because a reader told “not yet, 27 short” can come back and check, and a reader told “roughly two thirds” can do nothing at all.
What we refuse to measure
This is the part annual reports leave out. A reader who sees 6 good answers cannot tell whether we declined to ask an eighth question. So the refusals are published with the answers, and each one says not “it is not relevant” but what measuring it would push us to do.
- Time spent in the product, daily actives, sessions per week. Measuring attention is how a product starts competing for it. The Epic already refuses to reward speed; reporting engagement would reward it from the other side.
- Memories per traveler, as a success figure. The moment that number becomes a goal, the product starts extracting. It appears on this page only inside coverage, where what matters is the spread across a life and not the volume.
- Anything that needs reading what someone wrote. No number on this page is computed from the content of a memory, a conversation or a recording. The aggregates are counts and scores; nobody reads a sentence to make them.
- Satisfaction scores. We would be asking people to rate a product that holds their dead. The fidelity number already asks the only question worth asking, and it is harder to pass.
Where the numbers will come from
From the same aggregates the console reads: counts and scores, computed by functions that cannot return a word anyone wrote. No figure on this page is ever made by reading a memory, a conversation or a recording. The scale for fidelity is published separately and was fixed the same day — see the standard candle — and what an Echo may be asked, by whom, is the kernel.
Written 2026-10-04. First report 2027-01-31.