Identical inputs and seed produce identical results.
Transparent by design
Methodology
Every DeltaZero verdict is produced by documented calculations and deterministic thresholds. Numbers are computed in code; generative AI is not used to invent financial outputs.
Wallet analysis uses public data and requests no custody or signatures.
Reports identify the source, snapshot time, generation time, and data quality.
Stress inputs are constrained; sensitivity results are not price forecasts.
Governance record
Model card
DeltaZero is a transparent engineering risk model, not a statistically validated trading model. A version change is required when formulas, thresholds, factor distributions, or recommendation rules materially change.
Core model
Strategy metrics
Hedge ratio
short notional ÷ long notionalMeasures how much of the effective long exposure is offset by the short leg.
Hedge drift
|1 − hedge ratio| × 100Distance from a fully matched hedge, expressed as a percentage.
Net delta estimate
(long − short) ÷ long × 100A simplified directional exposure estimate for the submitted structure.
Net carry APY
long yield − weighted funding − fee dragFunding is weighted by the hedged share of long exposure.
Safety Buffer
min(100, collateral ÷ short × 200)A heuristic collateral-coverage score. It is not a venue liquidation price.
Capital at risk proxy
unhedged notional + margin deficitMargin deficit uses a minimum short-margin ratio of 10%.
Worked example
Why is the Safety Buffer 76?
min(100, $1,500 ÷ $3,950 × 200) = 75.95 ≈ 76The score expresses collateral coverage relative to short notional. The 200 multiplier maps 50% collateral coverage to the top of the 0–100 display range. It is a deliberately simple product heuristic—not a probability of safety, liquidation distance, or venue health factor.
Decision controls
Risk thresholds
| Profile | Target hedge | Drift warning | Drift critical | Buffer warning | Buffer critical | Provenance |
|---|---|---|---|---|---|---|
| Low risk | 0.92 | 4% | 8% | 70 | 50 | Conservative engineering policy |
| Medium risk | 0.96 | 6% | 12% | 60 | 40 | Baseline engineering policy |
| High risk | 0.98 | 8% | 16% | 50 | 35 | User-selected higher tolerance |
These thresholds are explicit product risk policies chosen for explainability and conservative intervention; they are not presented as empirically optimal or protocol-issued limits. Target styles apply their own allocation and intervention profiles. The most severe triggered condition takes priority.
Loss model
Stress impairment
Pre-stress equity is compared with equity after repricing the long leg, short PnL, collateral haircut, liabilities, exit slippage, liquidation penalty proxy, and protocol-loss assumption. The report displays each component separately to avoid hiding double counting.
Important
Safety Buffer and liquidation penalty are decision heuristics, not protocol-perfect liquidation engines. Venue rules, oracle behavior, fees, and latency may differ.
Sensitivity model
Monte Carlo
DeltaZero uses a correlated multivariate Student-t systemic layer for market movement, funding shifts, and collateral stress, then transmits collateral depeg severity into funding pressure, slippage, and collateral impairment. Non-systemic mode remains available as an independent normal baseline. Volatility scales with the square root of the selected horizon, and a supplied seed makes the simulation reproducible.
Path count
100–10,000Validated before execution.
Market bounds
−95% to +100%Both positive and negative shocks.
Reported tails
P95 and P99Alongside median and worst simulated outcome.
Repeatability
SeededSame request and seed, same paths.
Systemic tails
Student-t · 5 DoFConfigurable heavy-tail dependence across market, funding, and collateral factors.
Depeg transmission
CoupledCollateral depeg can amplify funding cost, exit slippage, and haircut loss in one path.
Current statistical assumptions
The submitted correlation matrix must be positive semidefinite. Default correlations are explicit scenario assumptions, not empirically calibrated forecasts. The model captures static heavy-tail dependence and depeg transmission, but not time-varying correlations, volatility clustering, order-book feedback, oracle latency, contagion between venues, or path-perfect liquidations. P95 and P99 remain percentiles of the submitted sensitivity model—not forecasts of real-world loss probability.
Evidence status
Validation record
Automated tests cover deterministic repeatability, percentile ordering, input bounds, recommendation states, and payment protection.
Versioned regression fixtures use frozen reference inputs produced by the documented formulas. These demonstrate behavior, but are not historical performance claims.
DeltaZero has not yet published a time-aligned historical replay with funding, liquidity, and margin rules frozen at each observation.
Thresholds have not been optimized against realized liquidations or portfolio losses. They remain transparent engineering policies.
Required before claiming validation
Publish a versioned replay dataset, remove look-ahead bias, include contemporaneous funding and venue rules, compare unhedged, original-hedge, and DeltaZero-adjusted structures, and report both favourable and adverse examples.
Reproducibility
What every result should disclose
Model version, selected risk profile, target style, and the recommendation policy used.
Seed, path count, horizon, distribution assumptions, bounds, and submitted stress parameters.
Provider, source snapshot timestamp, report timestamp, freshness, and data-quality status.
Triggered warning and critical thresholds, available metrics, and any unavailable or unsupported evidence.
Agent interface
Live Model Context Protocol server
Compatible agents can discover DeltaZero's typed tools and consume structured deterministic outputs directly—without endpoint-specific wrapper parsers.
Streamable HTTP endpoint
Production MCP
https://deltazero-production.up.railway.app/mcpInitialization, discovery, methodology resources, market context, and all calculation calls are temporarily free during listing review. The OKX Agent Payments Protocol implementation remains staged for later activation.
Tool inputs and structured outputs are generated from the same validated Pydantic contracts as the API.
MCP tools call the existing deterministic services directly; formulas and verdict logic are not duplicated.
Live market context plus Strategy Build, Hedge-Drift, Funding Stress, Monte Carlo, and the complete Risk Engine.
Agents can inspect capabilities and invoke tools freely during review; standardized payment enforcement can be restored with one deployment setting.
Data integrity
Sources and freshness
Public market and supported position data. Live responses carry a source timestamp and data-quality label.
Read-only lending data through configured Ethereum or Arbitrum RPC access.
Supported market and vault positions from Morpho's public API.
Manual yield, funding, collateral, fee, stress, and simulation inputs are labelled as user supplied rather than live market data.
Use responsibly
Limitations
- DeltaZero does not predict prices or profitability.
- Safety Buffer is a heuristic score and must not be interpreted as liquidation probability.
- Systemic correlations and tail parameters are scenario inputs; they are not yet calibrated to a versioned historical dataset and do not change dynamically by regime.
- Partial or unavailable integrations can produce incomplete portfolio coverage.
- Outputs are decision support, not financial advice or execution instructions.
- Users and agents should independently verify venue rules, liquidity, oracle behaviour, latency, and transaction costs.