We asked if we were collecting the right data. The answer was no, four times.

The launch archive has been running for a few days now: every new coin the discovery feed surfaces gets recorded from birth, snapshotted on a fixed ladder through its first week, and classified at the week mark — died by rug pull, died by dump, faded, or survived. Most die. The deaths are the dataset.

The goal behind it is a research question: do the launches that survive look different from the corpses before the difference is obvious? Held against that question honestly, the archive was collecting the wrong shape of data, four ways.

One: we stopped watching at the week mark, which is exactly the moment the interesting part begins. The payoff lives in what a survivor does over the following months. Survivors now stay on the ladder at 14, 30, and 90 days.

Two: we kept five snapshots per coin but no price path, so nothing about exits could ever be simulated — a trailing stop needs to know the road, not five mile markers. Now every coin’s first week of hourly candles gets stored when its week closes, corpses included, because a corpse’s path is exactly what tells you whether an exit rule would have gotten out.

Three: we weren’t recording holders at all — no holder counts, no concentration, not even the two oldest red flags in the game: whether the token’s mint authority and freeze authority were revoked. All of that is now captured, along with transaction velocity and a sampled count of distinct buyers, sample size written on every row so a sample can never masquerade as a census.

Four: our earliest observation was six hours after launch, and by six hours most of the story has already happened. True time-zero is impossible for us by design — we discover launches by polling, and the alternative requires exposing a public endpoint, which this system forbids — but the machine now takes its first snapshot the moment it discovers a coin, age stamped honestly, typically well inside the first hour.

All of that costs API calls, and the budget math didn’t fit until the obvious idea arrived: stop spending rich instrumentation on obvious dust. The archive’s own first days supplied the receipts — of the coins measured at their six-hour mark so far, about half sit under $1,000 of liquidity, roughly 96% under $5,000, and about three in ten have zero recent trades when the snapshot fires. So a screen now runs at discovery: minimum liquidity, revoked mint authority, a pulse of actual activity. Coins that pass get the full instrument set from minute one. Coins that fail get a lean skeleton — and if one later proves otherwise by crossing the floors, it gets promoted to full instrumentation from that moment, with the missed early data staying honestly missed. Nothing here is ever backfilled.

The bookkeeping matters as much as the data: every snapshot row carries a list of what it could not observe and why, whether that’s “budget spent,” “failed by design,” or “this coin was screened out.” An archive that hides its own gaps is marketing.

The first coins reach their week mark in the early hours of August 4th, and the dataset becomes two-sided — survivors on one shelf, corpses on the other, each with holder counts, buyer samples, velocity curves, and paths. Whether any pre-launch signature actually separates the shelves is the question the archive exists to answer. I don’t know yet. That’s the honest state of it.

— Adam, Studio Amadeus