Risk Methodology
Last updated: August 6, 2026
Every opportunity on Akashi carries a risk grade from A (safest in our catalogue) to E, backed by a 0–100 score where lower is safer. This page documents how that number is produced — engine v2 — and, just as important, what it does not claim.
The grade is Akashi’s own opinion. It is not a rating from a licensed agency, not a probability of loss, and not investment advice.
1. Three pillars
Each opportunity is scored 0–100 on three independent pillars:
- Protocol — the venue holding the deposit: how long it has run, its incident history, whether its program is upgradeable and by whom (an on-chain check of the upgrade authority; an off-curve PDA is treated as a multisig, an on-curve key as a single signer), and disclosed treasury or insurance backing.
- Asset — the stablecoin itself: what backs it, market capitalisation, on-chain liquidity, how far it has drifted from $1 (including the worst break in the last 180 days), and whether a freeze or mint authority still exists.
- Strategy — what the position actually does: plain lending, a leveraged loop, a yield-token derivative, an LP position, a vault delegating to a manager. Leverage, liquidation thresholds and pool depth all land here.
2. Weights depend on the strategy
A leveraged multiply position lives or dies on its mechanics; a yield-bearing stablecoin lives or dies on the coin. So the pillars are not weighted equally — the mix follows the category:
| Category | Protocol | Asset | Strategy |
|---|---|---|---|
| Lending | 45% | 25% | 30% |
| Vault (managed) | 25% | 15% | 60% |
| Yield-bearing stablecoin | 25% | 55% | 20% |
| Fixed yield / PT | 30% | 25% | 45% |
| Multiply (leveraged) | 30% | 20% | 50% |
| Liquidity pool | 35% | 40% | 25% |
| Anything else (default) | 35% | 35% | 30% |
3. How the pillars combine
- Sub-factors inside a pillar are combined with a weighted power mean (exponent 1.5) rather than a plain average, so one extreme problem is not diluted by several benign factors.
- A weakest-link floor then applies: the composite can never fall more than 15 points below the worst pillar. A single catastrophic component — an algorithmic asset, a single-key upgrade authority — cannot be averaged away.
- A constant chain floor of 5 reflects the Solana-level risk every position shares. It is a floor, not a discriminator: nobody scores 0.
- An opportunity built on top of another protocol inherits that protocol’s risk into its own Protocol pillar — composability is not free.
4. Live incidents override the model
The scored model above moves slowly. A fast layer sits on top of it and snaps the score when something is happening right now:
- An active depeg on the underlying stablecoin snaps the Asset pillar to the maximum.
- A TVL drop over 40% in 24 hours snaps the Strategy pillar to 100 — a run, not a rerate.
- An opportunity placed under review by an operator is floored at 70 (grade D or worse) until it is cleared.
Depeg detection is driven by an external stablecoin-ratings feed that refreshes four times an hour, and a risk recompute is chased immediately after it — so a live depeg reaches the published grade within roughly 17 minutes. See data sources.
5. Score to grade
| Grade | Score | Reads as |
|---|---|---|
| A | 0 – 22 | Blue-chip end of our catalogue |
| B | 23 – 35 | Solid, with identifiable trade-offs |
| C | 36 – 50 | Meaningful moving parts |
| D | 51 – 70 | High risk — leverage, thin assets, or weak controls |
| E | 71 – 100 | Most moving parts in the catalogue |
Every grade ships with its reasons in plain language on the opportunity page. If a grade cannot be explained, treat that as a bug and tell us.
6. Limitations — read this part
- The scale is relative, not absolute.Grades rank opportunities against each other within Akashi’s catalogue. An A is the safest thing here, not a safe thing.
- The score is not a probability.It is not an expected loss and does not translate to “X% chance of losing money”. An expected-loss model was designed and deliberately not built — the constants would need calibration against a failure history we do not have enough of.
- The constants are expert judgement. Weights and thresholds are informed by a curated history of 21 recorded DeFi incidents. That is enough to calibrate directions and classes, not per-protocol precision.
- Coverage is stablecoin-only.Opportunities outside the stablecoin catalogue are shown as “Not rated” rather than being given a number the model cannot support.
- Some inputs are simply undisclosed. Multisig thresholds, treasury size and insurance funds are published by only a minority of protocols. Where a fact is missing the model takes the cautious reading rather than assuming the best case.
- A grade is not a safety guarantee and cannot anticipate a novel exploit. Smart-contract risk is real at every venue on the board.
7. Version history
- v2 (current) — numeric 0–100 score and A–E grade over the three weighted pillars described above, with the live-incident layer.
- v1 (superseded) — a coarse Low / Medium / High tier taking the worst of the three pillars, with no numeric score. It is still computed internally for backwards compatibility but is not displayed anywhere, and it frequently disagrees with v2.
Scores and grades are recomputed hourly and again immediately after any live-incident signal. Only the current grade is stored — it is overwritten in place, so there is no grade history to inspect and a past grade cannot be reconstructed. The rate and TVL history behind it is retained; see data sources. Material changes to the model will be recorded in this section with the date.