Risk methodology
Every lending market, vault, token and chain 1delta serves carries a risk rating. This section documents how those ratings are produced: what is measured, what the thresholds are, and — just as importantly — what is deliberately not penalised.
The ratings are descriptive, not prescriptive. They exist so that an integrator can filter a universe of thousands of markets down to the ones that meet a stated bar, and so that a number shown next to a position can be explained.
One scale
Everything resolves to the same pair of fields:
| Field | Values | Meaning |
|---|---|---|
score | 1-5, integer | Severity. 1 is safest, 5 is worst. |
level | green, yellow, red | Band. green = score ≤ 2, yellow = 3, red ≥ 4. |
A score is not a probability and the steps are not linear. Read it as a bucket: 1-2 means nothing we screen for is wrong, 3 means something is unproven or uncertain, 4-5 means a concrete defect was found.
The API exposes the composed result on lending pairs and pools as risk
(maxTokenScore, configScore, and a breakdown array), and most listing
endpoints accept a maxRiskScore filter that defaults to 4.
The dimensions
Risk is scored independently per dimension and only then composed, so a reader can always see which of them produced a downgrade.
| Dimension | Unit rated | What it captures | Page |
|---|---|---|---|
| Protocol | lender (e.g. AAVE_V3) | Track record, audits, maturity, pool governance | Lending protocols |
| Governance | pool, market, token, vault | Who can unilaterally change or upgrade the thing holding the money | Governance |
| Oracle | priced market / feed | Whether the price a market liquidates on is sound | Oracles |
| Curation | curator, curated market | Who selects and parameterises isolated markets | Curation |
| Bad debt | market, reserve, vault | Losses that already happened, normalised by size | Bad debt |
| Vault | ERC-4626 style vault | Composite of curator, holdings, withdrawability, NAV integrity | Vaults |
| Asset | token, per chain | Peg behaviour, issuer/upgrade control, wrapper inheritance, liquidity | Assets and LSTs |
| Chain | chain id | Rollup decentralisation and exit guarantees | Chains |
Composition rules
Four rules decide how sub-scores combine. They are applied consistently across every pipeline.
1. Independent failure modes take the worst, not the average. Governance, oracle integrity and bad debt are separate ways to lose the same deposit. Averaging lets a sound timelock cancel out a broken price feed, so the level is the worst applicable dimension. Weighted means are used only inside a dimension — for example across a vault's holdings, where the weights are real economic exposure.
2. Observed failure dominates predicted safety. A market with two top-tier curators and live bad debt is rated on the bad debt. Audits, curator AUM and protocol age are proxies for expected safety; realised losses and confirmed misconfigurations are evidence.
3. Unknown is not safe, and not automatically dangerous either. A contract
the classifier could not resolve, a token nobody has listed, a market with no
curator — these cap at yellow. They never silently pass as green, and they
never alone force red. A gap in our coverage is not evidence against the
protocol.
4. Materiality gates before ratios. Every normalised metric divides by a size, and near-zero denominators manufacture infinite risk. Abandoned dust — a $9 remnant in an empty market, a test vault holding a dollar — is recorded as informational instead of rated catastrophic.
What is deliberately not penalised
- Volatility and RWA-ness. Lending against a volatile or real-world asset is ordinary lending risk, priced by LLTV. Only untrusted collateral (compromised, discontinued, or a proven broken peg) counts against a market.
- Illiquidity by design. Vaults with async redemption (LST unbonding, cooldown savings, epoch-based withdrawal) are not marked down for a withdrawal delay that is the product working as specified.
- Permissionless deployment. On isolated-market protocols anyone can deploy a market nobody uses. A single junk market does not paint the whole protocol red; the lender is rated on the share of its classified markets that are high-risk.
Where the data comes from
| Source | Provides |
|---|---|
| DefiLlama | Hack history with dated amounts, protocol TVL history, stablecoin supply and price history, audit links |
| Exponential.fi | Baseline protocol rating, auditor list, risk commentary |
| L2Beat | Rollup risk criteria (state validation, DA, exit window, sequencer/proposer failure) |
| Morpho Blue API | Curators, per-market curator index, market warnings, bad debt (realised and unrealised), vault allocations |
| Euler V3 Data API and entity registry | Vault governors, curators, Euler Earn allocations, lostAssets |
| On-chain reads (multi-RPC) | Pool and market admins, proxy patterns and upgrade authorities, mint authority, Aave getReserveDeficit, Gearbox credit-account solvency, Safe thresholds and owners |
| 1delta data layer | Market universe, collateral sets, token lists, prices, DEX liquidity depth |
| Nexus Mutual | Cover amounts and premium rates per protocol — a market-priced view of protocol risk, tracked alongside the scores but not folded into them |
| Frankfurter | Fiat reference rates for non-USD stablecoin pegs |
Curated inputs (asset whitelist, renowned-curator allowlist, rating overrides for protocols winding down, governance overrides for non-standard proxies) fill the gaps that no automated source covers. They only ever refine a specific entry, and each carries a stated reason.
Refresh and staleness
Hack data refreshes several times a day; the protocol, asset, chain, oracle, governance and vault pipelines run as one ordered sweep, each stage reading the previous stage's output. Two habits keep a stale input from reading as a clean bill of health:
- A snapshot older than the freshness window can only ever raise a vault's floor — never lower an existing red.
- A fetch that returns nothing fails loudly rather than overwriting good data with an empty file. An empty hack list would otherwise read as "no protocol was ever exploited".
Limitations
- Scores are point-in-time. A protocol that has never been exploited is not a protocol that cannot be.
- Coverage is uneven. Some markets are unscreened for oracle or governance risk;
those come back
unknownand cap at yellow rather than being assumed safe. - Share-of-markets is used where share-of-deposits would be the better denominator, because per-market TVL is not available for every family.
- Curator quality is measured by assets under management and identity verification, which are proxies for competence, not measures of it.