What we published, and what happened next.
Every liquidation cluster Isobath shows is written to an append-only record at the moment it is published, and the same bucket is measured again 6, 24 and 72 hours later. This page reports the state of that record. It does not report a success rate — and the reason it doesn't is the most useful thing on the page.
The record so far
Why there is no hit rate here
Recording makes no claim, so it needs no pre-registration. The moment anyone says "clusters we called vulnerable resolved N% of the time", that is a finding — and a finding needs a frozen contract with kill conditions, written before the outcome is computed. Publishing a number first and a method afterwards is how a record stops being evidence and starts being marketing.
There is also a concrete problem, found on day one of recording and measured again below: the headline variable barely moves. A predictor with almost no variance cannot discriminate, so the obvious question may be unanswerable as currently posed — not because the effect is absent, but because there is no contrast to measure it against.
How much each recorded variable actually varies
The spread across all 6,626 recorded clusters. Any future contract has to state which variable carries the contrast, and show it has enough spread to detect the effect it claims — before an outcome is computed. This table is what that argument will have to survive.
| Recorded variable | Min | Median | Max | Distinct values |
|---|---|---|---|---|
| Cluster size (USD) | $250,021 | $904,365 | $87,449,739 | 6,621 |
| Distance to spot (%) | 0.0% | 4.0% | 9.5% | 20 |
| Cannot-defend (%) | 0.0% | 100.0% | 100.0% | 638 |
| Thinly-defended (%) | 0.0% | 100.0% | 100.0% | 631 |
Cluster size varies across thousands of distinct values. The two "defended" percentages sit at or near their ceiling for most of the record — which is exactly the flatness that makes them unusable as a discriminator today.
What could confound this, stated in advance
Written down when recording began, not discovered later by someone defending a number.
- The scan set moves: a bucket can shrink because positions closed or because those wallets left the top 300. The raw data cannot tell those apart.
- Coverage moves between snapshots, so part of any change is a change in what could be seen.
- A fully liquidated wallet vanishes from the scan, which looks identical to closing voluntarily.
- Price path is sampled hourly, so whether spot reached a cluster is a lower bound.