VIGIL
Decentralized Liquidation Prediction
Bittensor Subnet Proposal
Decentralized Liquidation Prediction
Bittensor Subnet Proposal
liquidated in DeFi last year
😶 Borrowers had no warning
🤖 Liquidators with private bots won everything
The prediction intelligence exists. It's locked in private bots.
User deposits $100K collateral
Market drops 20%
Health factor crosses 1.0
LIQUIDATED
$10K penalty. No warning.
Miners compete to predict liquidations
Predict which positions will liquidate
Verify against on-chain events
Rewards accurate predictors
Bittensor network emissions — not users. Like Bitcoin mining rewards, TAO is emitted by the protocol itself and distributed to subnets based on value created.
Every hour:
Scoring:
| Dimension | Weight |
|---|---|
| Precision | 40% |
| Recall | 30% |
| Lead Time | 20% |
| Calibration | 10% |
Miner predicts liquidation. User adds collateral. No liquidation.
Was the prediction wrong? Not if HF dropped below 1.02.
→ Half credit for danger zone hits
Health Factor = liquidation risk metric.
HF > 1 = safe. HF < 1 = liquidatable.
→ Track HF over time via archive nodes
5-10 liquidations/day = too noisy for per-epoch scoring.
24-hour window = 50-200 data points.
→ Statistical significance
Earlier predictions earn bonus — but only if precision > 50%. Prevents wild early guesses.
Liquidations are on-chain facts.
| Data | Source | How |
|---|---|---|
| Did it liquidate? | Transaction hash | LiquidationCall event logs |
| When? | Block timestamp | Block header data |
| Danger zone hit? | HF history | getUserAccountData() on archive node |
Health metric values over time. Validators query each protocol's risk function (Aave/Morpho: getUserAccountData, Compound: getAccountLiquidity) at historical blocks. If the position entered danger zone, we can prove it.
No subjectivity. No disputes.
Validators verify facts, not judge quality.
| Customer | Product | Why They Pay |
|---|---|---|
| Borrowers | Alerts | Avoid 5-15% penalty |
| Liquidators | API | Better timing |
| Protocols | Dashboard | Risk monitoring |
Prediction ≠ Execution
• Prediction requires analysis skills and data infrastructure
• Liquidation requires $100K+ capital and MEV infrastructure
Vigil commoditizes prediction. Liquidators compete on execution.
vs SN10 (Sturdy)
| Sturdy | Vigil |
|---|---|
| Maximize yield (offense) | Predict risk (defense) |
| Continuous optimization | Discrete events |
| Weeks to verify | Hours to verify |
vs Outside Bittensor
| Competitor | Gap |
|---|---|
| DeFi Saver | Reactive, not predictive |
| Gauntlet | Centralized, B2B only |
| Private bots | Proprietary |
Vigil = First decentralized, predictive liquidation intelligence
4 weeks. Protocol-agnostic architecture. Aave + Compound + Morpho = 65% of lending market.
Adapter interface
Aave V3 adapter
Compound V3 adapter
Morpho adapter
Validator scoring
(works across all)
Testing + docs
Historical replay
Each protocol defines: prediction window, risk metric, liquidation events. Scoring logic is protocol-agnostic. Add new protocols by implementing one interface.
Maker uses auctions (not instant). Architecture supports 24h window + auction-start prediction. Documented, not built in Round II.
Decentralized Liquidation Prediction
Key Innovations:
Faruukku