✦ The risk engine
Knows when to step back.
Most platforms keep trading no matter what. Ours doesn't. The risk engine reads the market every few seconds, spots crashes 60 to 180 seconds before they hit, and pauses every strategy when the signal fires. The defense isn't optional.
Regime classification
Direction
TRENDING
Violence
MODERATE
Cost of lev.
NEUTRAL
Activity
STRONG
Cascade Guard
SCANNING FOR CRASH SIGNALS...
Signal strength
4
Market checks
60-180s
Crash warning lead
Continuous
Market read
Auto-pause
On crash signal
✦ The platform's job
The hardest discipline in crypto.
The hard part of algorithmic trading isn't picking strategies in good markets. It's not getting caught on the wrong side of a LUNA collapse, a Black Thursday liquidation spiral, an exchange outage during a 30% move.
Most platforms have no risk engine at all. They execute their strategies regardless of conditions, on the theory that the user will turn them off in time. Users don't. Our risk engine sits between the strategies and the market. When conditions get bad, the engine pauses. When the market changes character, the engine rebalances. The discipline you wouldn't maintain manually is the default.
The platform's job is to know when not to trade. Most don't have one. Ours does.
✦ The 4-factor regime classifier
Four checks. One vote. Continuous.
The classifier reads four signals continuously and votes on the current market state. Each one is computed independently from live data. The vote decides which strategies run, which sit out, and at what weighting. When three of four disagree, exposure shrinks. When the market flips character, the strategy mix rebalances.
Direction.
Where is the market moving? Trending, ranging, or chopping. Drives the mix between trend-following and mean-reversion strategies.
Violence.
How aggressive is the market right now? Calm, elevated, or extreme. Drives how big each trade can be and how tight the stops are set.
Cost of leverage.
What are other traders paying to be on each side? Reveals where the crowd is positioned. Drives the windows when funding-rate strategies can earn yield.
Activity.
How busy is the market? Real participation versus thin liquidity. When markets get thin, execution quality drops and so does our exposure.
Four signals. One vote. Updated continuously.
✦ Cascade Guard
The 60-second early warning.
Crashes are the dominant tail risk in leveraged crypto. A large position gets force-liquidated, the price moves, that triggers more liquidations, the crash propagates across venues, prices move 10-30% in minutes. By the time a human notices, the move is over.
Our crash detector reads on-chain liquidation data in real time across major venues, models the likelihood of a propagating crash, and detects them 60 to 180 seconds before they hit the broader market. When the signal fires, every strategy stops opening new positions on its own. Existing positions are managed according to each strategy's own rules — this isn't a stop-loss, it's a "don't add fuel to the fire" rule.
The crash comes. The detector fires. Strategies stop. You don't have to be paying attention.
✦ The fee floor
The 82%-win, 0%-profit defense.
A strategy with an 82% gross win rate that returns 0% net of fees is a strategy losing money on every trade it wins. This was an internal Perpblock failure mode early on — beautiful backtests, terrible live results — and the fee floor is the structural fix.
Every profit target on every strategy must clear the venue round-trip fee by 50% before it deploys. Below the floor, the strategy refuses to start. The check runs at strategy startup, not silently during execution. If a strategy's targets sit below the floor, the deployment fails with an explicit error showing which target is too tight and what the floor is.
Strategies are validated before they trade. Not after they lose money.
✦ Risk profile and guardrails
Your stated profile drives every decision.
Onboarding includes a risk quiz that produces a real profile — tier, capital range, drawdown tolerance, time horizon, experience level. The profile drives concrete decisions across the platform. Strategy selection respects your tier. Allocation weighting respects your drawdown tolerance.
Beyond static configuration, a continuous watchdog checks your portfolio against your stated profile. When the portfolio drifts out of profile — concentration creeps up, drawdown crosses a level — a warning surfaces and the AI references it when it next talks to you. Your profile isn't a one-time form. It's the input the platform consults on every consequential decision.
The platform behaves the way you said you wanted it to behave.
Strategy tier: Moderate ✓
Drawdown: within −18% tolerance ✓
Concentration: 34% BTC — elevated ⚠
Leverage: 3× · acknowledged ✓
Watchdog status
✦ Leverage gating
Leverage is a license. You earn it by passing the quiz.
Leverage tiers above 1× are gated behind a knowledge quiz. Acknowledging 3× doesn't unlock 5×. Acknowledging 5× doesn't unlock 10×. Each tier requires a separate quiz pass.
The leverage advisor — a deterministic system, no AI involved — ranks every leverage level as recommended, acceptable, risky, or blocked for your profile and the specific strategy. If a level is blocked because you haven't passed the quiz, the UI shows "blocked — quiz required" with a link to the quiz. Profit targets, stop levels, and trailing stops are calculated against the fee tier and shown before you commit.
Leverage isn't a slider. It's a license. You earn it by passing the quiz.
✦ The policy layer
Rules decide. The AI explains.
Asking an LLM "should you push back on this?" is unreliable in both directions. We use a deterministic policy system instead. Rules are written, testable, and version-controlled.
Examples of what the policy system checks: does this action exceed your tier allowance? Does it break your concentration cap? Does it conflict with a goal you told the AI earlier? Is it inappropriate for the current market read? The evaluator runs before any consequential action — strategy enrollment, allocation change, manual trade — and returns a verdict of pass, warn, or block. When it returns warn or block, the AI explains the objection in plain English. The rule logic is auditable. The prose isn't where the decision lives.
Rules decide. The AI explains the decision.
Verdict: WARN · BTC concentration at 34% exceeds 30% cap. Rebalance before increasing exposure.
BLOCKED: BTC-TREND failed 3 of last 5 trades. Cooling-off period: 47min remaining.
✦ Built to refuse
We'd rather block a trade than apologize for one.
Before any consequential action — strategy start, allocation change, leverage change, large position open — the platform runs a sanity check. The check covers four things at once: is there enough market depth for the trade size you're proposing, does the action satisfy the fee floor, is the position size consistent with your profile, and has this strategy failed recently in a way that should pause new commitments?
Failed sanity checks return a structured refusal that the user can read and the AI can explain.
The platform refuses bad actions. We'd rather block a trade than apologize for one.
✦ FAQ
Common questions.
✦ Defense first. Strategy second.
Most platforms ship strategies and call the defense optional. We shipped the defense first.
Crash detector on by default. Fee floor enforced. Risk profile respected.