Incident Archive · Cascade Forensics
Selected cascades, reconstructed from the archive.
Complete archive · 353 L3-classified cascade events (distinct events, internal rule) across all 47 markets · $2.376B liquidation notional · 100 events carry a full replay · archive as of 20 August 2026 (replay set revised 5 October 2026).
Every replay is an event classified by RiskLoom's internal L3 cascade rule, reconstructed minute by minute from stored records — 2026 production data, real-notional, single methodology. CRI is an empirically calibrated structural stress index derived from perpetual market behaviour. Each replay shows when CRI crossed its 0.495 threshold, when the cascade started and when its liquidations peaked: the sequence as recorded, not a contemporaneous alert.
93.2%
Notional coverage · 329 of 353
Every figure above describes the
full archive as of 20 August 2026 — not the replay subset. The archive is live and grows;
this page is a stamped snapshot of it.
Archive methodology
Every row is a real 2026 leverage cascade, identified from the liquidation record and
admitted only if CRI crossed 0.495 within the 120 minutes before its start. Liquidation notional is atomic — the sum of five-minute
liquidation buckets inside each cascade window, single methodology, no rolling-aggregate
inflation. It resolves for 329 of 353 events
(93.2%). Of the remaining 24, 2 are priced on the reconciled basis (†) and 22 are shown as
“—” with every other field intact: no canonical five-minute liquidation record covered
their window at the snapshot.
Correction — one cascade was being counted many times
Until 19 August 2026 this archive counted 475 rows where there were
353 distinct cascades. The detector emits a new row on each cycle while
a cascade is still extending — the same event, its start marching forward in five-minute
steps against a fixed end, each row capturing slightly less of the same window. That
inflated the count by 122 rows (26%).
One figure we previously published was wrong because of it. The overlap rate was
stated as 41% of cascades sharing a minute with another; almost all of that was a cascade
overlapping itself, not genuine cross-cascade overlap. Measured on distinct events it
is 40 of 353
(11%).
The liquidation total did not move: it was already deduplicated over distinct
minutes, so the double-counting never reached the dollar figure.
Why the archive total is smaller than the sum of its rows
Detection windows can still genuinely overlap: 40 of 353
cascades (11%) share at least one minute with a
different cascade in the same market. The archive total published above is therefore
deduplicated over distinct liquidation minutes — every liquidated dollar counted
once, $2.376B across 3,799 distinct minutes.
Adding up the per-event figures in the table gives a larger number
($2.510B, 6% higher),
because minutes shared by overlapping windows would be counted more than once.
Both figures are correct for what they measure: each row states the notional inside
its own cascade window, and the archive total states what was actually liquidated.
Non-events, misses and the full signal record are reviewed during evaluation-stage
methodology sessions.
Liquidated side is which side was forced out, not a
trade direction: 232 cascades liquidated longs (forced selling),
118 liquidated shorts (forced buying). Venue co-occurrence means the event also appears in Hyperliquid
liquidation research data (264 of 353); it is not independent confirmation.
Featured replays
The 12 largest documented cascades · click any card to replay it step by step
The full archive
—