What every score means
All scores run 0–100 where 100 = safest. Grades: A ≥85 · B ≥70 · C ≥55 · D ≥40 · E <40.
Since 5 October 2026 the grade is the weighted average of eight factors: the exchange's insurance fund 15% · assets vs liabilities 15% · insurance availability 15% · sentiment (news plus the move in its token or share price) 15% · transparency 10% · operational 10% · regulatory 10% · market presence 10%. Each factor's method and sources are shown when you click its box on the dashboard. A factor that cannot be measured is left out and the grade is capped instead (at B if one of the first three is missing; at C if under 60% of the grade is measured). An insurance fund that falls 20% below its 7-day high scores zero. Holdings of an exchange's own token never count as a fund. The categories below are the underlying measurements that feed those factors.
TVL stability asks "is capital fleeing?" — it measures 30-day drawdown and volatility and is size-blind by design: a steady $0.5B venue beats a wobbling $5B venue here.
Liquidity / activity (DEX) and Loss absorption (CEX) ask "how big is the cushion?" — absolute depth and balance-sheet scale, log-scaled so each 10× in size earns the same increment.
One detects stress, the others measure capacity. A venue can score high on one and low on the other — that is intended.
CEX categories (asset-side only — liabilities unobservable)
Reserve adequacy weight 5%
How many times do on-chain reserves cover current open interest (the venue's live derivative commitments)? Coverage ≥2× is strong; <1× is a concern. Where DefiLlama has no reserve data, a BTC-only CoinGlass estimate is used (capped 85); with no data at all the floor is 40 (55 with third-party-attested PoR).
Ignores: absolute size (that's loss absorption) and customer deposits beyond open positions.
Flow stress weight 20%
Is money leaving faster than normal? BTC balance changes over 7/30 days (outflow shared with the cross-exchange median counts half — sector rotation isn't venue stress), liquidations vs OI, and funding-rate dislocation vs peers. The most direct early warning of a bank-run.
Reserve quality weight 10%
What the reserves are made of. Own-token share is the FTT/FTX pattern: an exchange whose "assets" are its own printed token can collapse overnight. Penalised only once it is a dominant share — no deduction below 45% of reserves, then −1.8pts per excess point (a near-all-own-token book bottoms near zero). A 10–15% holding (e.g. BNB) is no longer marked down here — the own-token crash failsafe still watches its price live, and loss_absorption already scores on clean reserves. Above the 45% line it also penalises an own token sliding >20pts vs BTC over 30d. Low weight: a latent tripwire that's flat at 100 for the current board.
Unobservable composition: 100 if the venue has no native token AND an attested PoR (verified clean — don't dock a transparent venue for a feed gap), 90 if no native token alone, 55 with an attested PoR only (below an honest disclosure), else 50.
Transparency weight 20%
How verifiable is the venue? PoR published (+30) and live (+20), update cadence (+20), independent attestation or zk methodology (+20), proof of liabilities (+5), machine-readable status feed (+5), independent tracking by CoinGlass/CoinGecko (+10 each).
Loss absorption weight 15% — added 2026-06
Black-swan capacity: the absolute balance sheet available to make users whole after a major incident, the way Binance's SAFU fully covered its 2019 hack. Measured on clean reserves (own-token holdings excluded — a cushion of the venue's own printable token is exactly what failed at FTX). Log scale: $1B → 40, $10B → 65, $100B → 90. Unknown reserves score a neutral 50.
The CAPACITY axis — distinct from insurance (willingness) and reserve adequacy (a coverage ratio).
Regulatory weight 10% — asset-safety lens
Enforcement exposure, scored for "will I lose my parked assets" rather than for past conduct:
100 − penalty × factor × 3. An active compliance monitor/probation is treated as mitigating
(factor ×0.5) — the venue is resolved, supervised and cleaned up, and the underlying actions are mostly AML/sanctions
(illicit-finance controls), a weak proxy for customer-asset theft. The real asset-safety risk is an
unsupervised grey-zone venue (unregistered, no regulator catching problems early — the FTX setup),
which takes the full penalty. A resolved, unmonitored matter decays 5-year linear (floor 0.25).
Insurance weight 15%
Two blended axes (60/40). Willingness: when users actually suffered losses, did the venue make them whole? — a researched payout track record (strong 90 · mixed 60 · poor 35 · untested 65), where "a big fund ≠ they'll pay you." Client insurability: can a client buy third-party cover on assets held here, and what does the market charge? — the live Nexus Mutual custody-cover premium, inverted (a cheaper premium means underwriters see the venue as safer); no active cover → 40. The premium is a market-priced risk read; the available capacity is single-digit-$M, so this is a signal, not a full hedge for size.
Distinct from loss_absorption / vault_counterparty (capacity — how much they CAN cover): this is whether they WILL pay and whether the market will insure you. Fund size is not re-scored here.
Operational weight 5% — regulatory split out 2026-06
Incident history (log-severity, 18-month decay), live status-page incidents, and a longevity bonus (+2pts per year of operation, capped +10). Regulatory enforcement is now its own factor (above).
DEX categories (reserves observable on-chain; contract/vault risks apply)
Contract / incident weight 25%
The existential risk for a DEX. Audit status (audited = 80 baseline, unaudited = 45), exploit history (severity-weighted, 18-month decay), and a mainnet longevity bonus (+2pts/year, capped +10). Absorbed the retired infra-concentration weight.
Vault counterparty weight 10%
The loss-socialisation buffer. When the venue's own LP vault absorbs liquidations, one toxic position can transfer losses to depositors (the JELLY vector). Scored on adequacy (backstop vs open interest, anchored to industry backstop sizing: 1% of OI ≈ 53, 3% ≈ 69, ≥5% caps 85) plus the vault's own 30d drawdown trend. Where a venue runs a secondary assistance fund denominated in its own token (Hyperliquid's, valued live from the venue's public API), it counts toward adequacy at a 50% haircut — real in moderate stress (it covered the JELLY loss) but reflexive in venue failure: the token crashes exactly when the fund is needed. No vault → 70 neutral; a vault that exists but can't be independently measured → 50 — the system-wide "unknown" price (opacity is itself a risk fact); a publishing vault whose feed gaps on our side → 55.
Liquidity / activity weight 35% — top DEX factor
Exit capacity: can a desk get out under stress? Absolute TVL and market-cap/OI depth on a log scale (0pts at $10M → 20pts at $10B per component, base 60). This is where venue size earns credit, and it carries the most weight on the DEX side — for a trading desk, depth and exit capacity are the dominant counterparty concern: no desk sizes a $0.5B venue like a $6B one.
TVL stability weight 5%
Capital-flight detector: 30-day drawdown (−2pts per 1%) plus volatility (−1pt per 1% CV). Size-blind by design — it measures whether capital is leaving, not how much capital there is.
Infra concentration monitored, unscored — retired from the composite 2026-06
Single-point-of-failure risk. Rollup venues use the L2Beat stage rating (Stage 2 = 90, Stage 1 = 75, Stage 0 = 50 — "training wheels": users depend on the operator for exits); otherwise the top chain's share of TVL. Still computed and displayed on the venue page and fed to the AI risk manager as context, but carries zero weight in the composite (team decision): it is the crudest signal on the board, and the exit-capacity concern it proxies is measured better by liquidity/activity.
Regulatory weight 10% — added 2026-06
Same standalone formula as CEX. For DEX venues it captures grey-zone / unregistered / enforcement exposure — e.g. an FCA unauthorised-firm warning, or a US-incorporated operator running leveraged perps without registration. Scored as an ongoing posture (full penalty, no decay) until a venue's legal status resolves.
Insurance / payout weight 10% — added 2026-06
Payout track record (willingness axis), symmetric with CEX: did the venue compensate users in real loss events? A discretionary make-whole (e.g. Hyperliquid's JELLY settlement, Lighter's outage refunds) reads better than silently socialising losses to vault depositors. Distinct from vault_counterparty, which scores backstop capacity.
Always-on failsafes (independent of the AI layer)
- Composite drop ≥15pts in one run (critical ≥20).
- Solvency floor — reserve_adequacy / flow_stress / vault_counterparty under 20 (critical) or eroding under 30.
- BTC outflow — ≥5% over 30d, or fast + persistent 7d outflow; market-wide moves half-weighted.
- Own-token crash — venue token down ≥10% in 24h (warning); critical when it drops ≥7% while the peer exchange-token median doesn't (idiosyncratic = the FTT pattern; BNB/OKB/BMEX/HYPE/LIT all tracked).
- TVL/AUM velocity — venue TVL/AUM down ≥10% in 24h (warning); critical when down ≥5% AND ≥5pts below the cross-venue median — one venue bleeding AUM while peers are flat is an exodus (the FTX reserve-drain pattern), not the market.
- Venue dark — a venue that fails to collect/score posts a warning: silence is itself a risk signal.
- Ops health — heartbeats post if the AI layer goes silent or the news feed dies.
The unobservable liability side
All CEX scores are asset-side only. A centralised exchange's customer liabilities (deposits owed, derivatives exposure) are not independently observable from outside, regardless of Proof-of-Reserves quality — PoR attests what a venue holds, never what it owes. So no CEX score, however high, implies verified solvency; it reflects strong observable asset-side indicators only. DEX reserves are directly observable on-chain, but smart-contract, vault-socialisation and validator-centralisation risks apply in their place.