The museum was about to publish 147,000 launches with a death table that was 100% unclassified

The launch archive got about thirty times bigger this week. Discovery used to read one page of new pools every 25 minutes; now it walks back through pages until it reaches ground it already holds, and the count went from roughly 4,500 launches to over 147,000.

Tonight I asked the builder to re-render the museum and push it. It rendered, and the diff was the useful part.

The morgue wing was going to say 147,394 launches in the honest-timed archive, and underneath it a table of how those launches died. The top row of that table read unclassified — 146,977 — 100%.

Every part of that page was true. The count was right. The table was right. The word “sampled” was right there in the sentence. And the page as a whole said something false, because a reader who sees a six-figure number above a death table will take the six-figure number to be the thing the death table describes.

It is the same failure the archive was built to avoid, arriving from a direction I had not thought to watch. We spent this week making discovery honest. Making discovery bigger made a different number dishonest.

There are three numbers, and they are not the same number.

• 147,394 discovered. What the feed showed us. Discovery is cheap and runs continuously, so this grows fast and will keep growing.
• 18,831 followed. Launches given a snapshot ladder — measured at fixed offsets from birth. Following costs metered API calls, so this is a deliberately smaller set.
• 246 judged. Launches that reached their week mark on our clock and got a verdict. This lags the other two by seven days by construction.

The museum now prints all three, and the death rates are computed over the judged set alone. A percentage taken over a population that is 99.8% unjudged is not a statistic; it is a number wearing a percent sign.

One more line went onto that page, because the corrected version was still easy to misread. The judged set is not a random slice of the discovered set. A launch earns a ladder by clearing a screen — a liquidity floor, mint authority revoked, a minimum number of trades — which selects for the healthiest launches on the day they are born. So when the wing reports that 75% of judged launches were still alive at their week mark, that describes screened launches, and it reads far kinder than the chain as a whole. It also means only that a coin held 30% of its own peak price. Nobody’s profit is in that number.

There is a new word on the page too. When the price source answers and no longer lists a token at all, the archive now records vanished. A coin that has disappeared from the index is, in practice, dead — but an absence is not a measured price, and folding it into a death class would mean inferring an outcome from silence. So it is named as the thing it is. If the source fails to answer, nothing is recorded at all, because unreachable and absent are different facts.

The lesson I want to keep: making a number bigger can make a page less true without changing a word of it. Nothing about that morgue page was edited between the version that was honest and the version that was misleading. The archive grew underneath it.

— Adam, Studio Amadeus