19 claims · 2 refuted by us · none removed

Our record

Every claim Isobath has published, and what happened to it. Refuted findings are never removed — a record that keeps only its wins is a marketing page.

• MEASURED  Free collateral cannot be derived from position and margin data; naive arithmetic misclassifies one wallet in five
  withdrawable == accountValue - totalMarginUsed for 19% of wallets; 20% misclassified as able to defend
• MEASURED  Order-book depth falls as price moves get larger, so a static-book cascade estimate is optimistic
  near-book depth falls to 0.8462 at a 5-minute horizon during the largest moves (p25 0.6573), against 1.0015 in quiet markets
• MEASURED  A stratified estimate puts the discovered wallet universe at 53.3% of Hyperliquid BTC open interest, not 20%
  estimated 53.3% full-universe coverage, 95% CI [40.9%, 70.8%]; this is not the coverage observed in a live scan
• REFUTED  The wallet universe is exhausted; more wallets of the same kind would add nothing
  claimed 20.24% observed coverage was near the ceiling; a later stratified estimate put the wider discovered universe at 53.3% [40.9%, 70.8%]
• MEASURED  Daily-horizon directional tests on one instrument cannot detect realistic edges; the region is unpowerable, not empty
  detecting a 52% hit rate needs 4,900 independent days (13.4 years); 51% needs 19,600 (53.7 years)
• ASSUMED  Binance order-book behaviour under stress transfers to Hyperliquid
  none — nothing has been measured on Hyperliquid
• REFUTED  No provider sells per-wallet account state, so an archive of it is a durable moat
  HyperTracker sells it — complete market coverage, 15-20 minute snapshots, history from 2025-04-04, $179-1,999/month
• MEASURED  The position map is up to an hour stale relative to the order book, which manufactures phantom triggers
  clock gap of 3,055 s (51 min) between the latest l2Book snapshot and the latest position snapshot
• PRELIMINARY  Visible forced flow on Hyperliquid BTC is small relative to visible book depth
  $63.6M total forced selling in the visible map against $222.5M of bids within 2.71% of mid; a $10.6M trigger moved price 0.069%
• MEASURED  In CASCADE-1, forced-liquidation episodes did not move price more than volatility-matched random minutes in the same hour
  15m mean +40.07 bps in the forced direction against a matched control of 44.52 bps; no horizon clears both benchmarks
• PRELIMINARY  The hourly map systematically under-represents the positions closest to liquidation
  deep scan (5,395 wallets): median distance to liquidation 15.5%, 42.8% of positions within 10%. Fast scan (top 300 by notional) at the same period: median 31.6%, 23.0% within 10%.
• MEASURED  The vulnerability figure Isobath publishes is very nearly a constant, so it cannot discriminate between clusters
  across 433 published clusters, cannot_defend_pct has median 100.0 and 71% sit at exactly 100%. Splitting at the median leaves 29% below and 0% above. thinly_defended_pct behaves the same: median 100.0, 27% below, 0% above.
• PRELIMINARY  In four scan transitions, wallets holding free collateral disappeared more often; disappearance cannot be read as liquidation
  across 4 deep-to-deep transitions, of wallets within 5% of liquidation: 62.7% of those WITH free collateral were absent at the next scan, against 45.1% of those with none. The sign is opposite to the hypothesis that free collateral aids survival.
• PRELIMINARY  Both venues' ±1% BTC book held roughly a quarter of a billion dollars; measured retail-size sweeps cost about 1 bp and typical clusters were small relative to standing depth
  standing notional within ±1% of mid is $247.6M median on Binance futures and $225.6M median on Hyperliquid — within 9% of each other. At burst level, a same-side sweep worth 0.5–1% of that book costs the taker 0.8–1.1 bps of slippage. The median published liquidation cluster is 0.44% of the Hyperliquid book; the p90 cluster is 5.1%.
• PRELIMINARY  How a liquidation price responds to added margin cannot be derived from Hyperliquid's published state, so the "add $500 and your liquidation moves to X" feature cannot be built from arithmetic
  the documented formula reproduces the venue's own liquidationPx to within 10 bps for only 56.4% of cross positions (median error 4.79 bps, n=1,019); inverting it position-by-position within a single account recovers margin_available values that disagree by 3.49% at the median and 73.7% at p90, so only 28.2% of accounts are internally consistent within 1%
• PRELIMINARY  Across six years, none of nine public market-state variables separates forward trading conditions once trailing realised range is held constant
  band-controlled lift for all nine candidates sits between 0.914 and 1.129 in raw form, and z-scoring each against its own 30-day history collapses every one toward 1.000 (1.129 to 1.039, 1.126 to 1.052, 0.914 to 0.964). The raw long/short lift reverses sign by year: 1.170, 0.990, 0.948, 0.869, 1.164, 1.232, 1.178 for 2020 through 2026
• PRELIMINARY  In a uniform sample of discovered BTC-trading wallets, ten wallets held 57% of sampled leveraged exposure and the median position was $1,308
  of 8,000 wallets drawn uniformly at random from 51,765 discovered, 2,249 held a BTC position totalling $222.9M. The top 1 wallet holds 14.0%, the top 10 hold 56.6%, the top 100 hold 87.6%. Median position $1,308 against a mean of $99,108 — a mean 76 times the median
• PRELIMINARY  The hourly scan reproduces coverage and forced notional to within a point, but sees only 28% of positions near liquidation where the full scan sees 47%
  across 7 deep scans paired with the fast scan nearest in time, coverage agrees to within 1.6 points (36.8% vs 36.2% on the closest pair) and forced sell-side notional within ±12%; but median distance to liquidation is 11.6% in the deep tier against 26.5% in the fast tier, and the share of positions sitting within 10% of liquidation is 47.4% deep against 27.8% fast
• PRELIMINARY  Holding trade size against the standing book fixed, a book thinner than three-quarters of its own daily normal costs about 20% more — so the ratio absorbs most but not all of the stress regime
  across 985,950 aggressive bursts, median cost at a burst/depth ratio of 0.01-0.02 is 1.88 bps when the book is below 0.75x its day's median, against 1.62, 1.54 and 1.62 bps in the three deeper regimes. The pattern repeats at every ratio band: 0.98 vs 0.80-0.83 at 0.005+, 3.24 vs 2.58-2.74 at 0.02+. Only the thinnest band separates; the other three are indistinguishable

Every finding is also a machine-readable object: data/findings.json carries all 19 with their status, sample, method and stated limits.