Trustless Harvester™ — Technical Blueprint

01What it is, and why it's different

The Trustless Harvester is an autonomous agent that provides concentrated liquidity on a DEX to earn trading fees — but only while the market is calm, and fully in stablecoins otherwise.

Concentrated-liquidity managers (CLMs) are not new. Beefy, Gamma, ICHI, Range, Steer and others all automate LP positions. Their shared design choice is to stay always deployed: liquidity is the product, so the pool is never empty. That is exactly why they track the market down — an always-on LP eats the underlying's full drawdown to keep earning fees.

The Harvester makes the opposite choice. Being in the pool is a decision the engine takes each cycle, not a standing default. It harvests fees in calm regimes and steps entirely aside to USDT when the flow signals a turn. Same fee source, inverted risk posture.

02Where the yield comes from

Every swap that trades through your active price range pays the pool's fee tier, split pro-rata across the liquidity that is in-range at that tick. The yield is real cash-flow from trading volume — not emissions, not a token subsidy.

The tighter the band around spot, the larger your share of that fee for each dollar of liquidity. That is the whole reason to run a narrow, spot-straddling range — and also the whole reason not to hold it into a move, since a tight range is the first thing a trend walks out of.

03The control loop

One agent runs a continuous sub-second loop. It is the spine everything else hangs off:

Every pass — deploy, hold, exit, or skip — is written to the log with its full market context, so the same decision can be replayed and the model recalibrated against what actually happened next.

04Where the data comes from

The engine reads four independent streams. No single source can be spoofed into a false regime, because they have to agree.

05Signal layer

From those streams the engine derives ~70,000 signals refreshed every 500 ms. The features are deliberately flow-based, not price-based:

• CVD (cumulative volume delta) — net aggressive buying vs selling.
• OFI (order-flow imbalance) — pressure building in the book before it prints.
• Book depth & slope — how much size it takes to move price.
• Cross-asset correlation — is this pair moving with the macro, or alone.

Each feature is normalized to a rolling baseline (z-score), so the engine responds to change in pressure, not to absolute levels that drift over time.

06Regime classification — the calm/storm decision

A multi-factor state machine fuses trend, realized volatility, order-flow direction, and breadth into a small set of regimes:

• Hysteresis + persistence. A regime must confirm across consecutive windows before it flips — no deploying on a one-bar calm, no exiting on a single noisy tick.
• Fail-closed. Any data gap, stale feed, or ambiguous state resolves to no exposure. The default is cash. Wrong-and-out costs a spread; wrong-and-in costs the drawdown.

07Where capital goes — venues & lifecycle

Deposited USDT lives in the vault and moves along one deterministic path. Between deployments it sits idle in stables — the out-of-market default.

08Liquidity placement

In a calm regime the keeper opens a narrow concentrated range directly on PancakeSwap V3 — no rented rebalancer, no third-party manager contract in the fund-path. The band is placed to straddle live spot, because a range price never traverses earns nothing. Width is sized to the calm-regime volatility the classifier expects: wide enough to survive normal noise in-range, tight enough to keep fee density high.

09Inventory & impermanent-loss posture

IL is not hedged away — it is avoided by discipline. Because the engine is out of the pool during directional regimes and only runs tight spot-straddling ranges in calm, it never carries a large open inventory into a trend.

10The exit

On regime degradation the position is fully unwound to USDT — not rebalanced, not re-centered. A single decisive de-risk, ahead of the move the flow predicted. Entry and exit run off the same signal layer, so they share one coherent view of the market instead of two rulesets that disagree at the worst moment.

11Vault architecture — non-custodial

Deposits sit in a standard ERC-4626 vault that mints shares pro-rata. Share price = NAV per share, computed on-chain from vault holdings marked to market. Withdrawals are hard-wired to the depositor: redeeming burns shares and returns assets at current NAV. The right to exit is a contract guarantee, independent of the operator or the interface.

12Trust model — keeper vs contract

The keeper (the agent's on-chain signer) holds trade-only authority: route swaps and manage liquidity positions inside the vault, nothing else. It cannot withdraw, cannot change ownership, cannot redirect funds. Segregation of duties is enforced by the contract, not by policy — there is no admin path that reaches principal. "Nothing to rug" is a property of the permission set, not a promise.

13NAV accounting & fees

NAV is marked continuously from pool positions plus idle stables. The 20% performance fee is measured against a high-watermark and crystallized only on redemption, only on gains above the prior high. No management fee, nothing on principal. Flat / down / sideways ⇒ fee = 0. Operator and depositor share one number: NAV.

14Verification & security

• Symbolic proof (Halmos) — withdrawal-safety and no-keeper-drain invariants are proven, not just tested.
• Static analysis — Aderyn + Slither across the contract surface.
• Forge suite — 332 / 333 property, unit, and invariant-fuzz tests.
• 5 adversarial review rounds on the contract set.

The claim is specific: the depositor can always redeem at NAV, and no operator action can take that away.

15Data & calibration

Every decision is logged with full market context (141 fields per trade) into a time-series + event store: 574M+ points, 8M+ evaluations, 340 days of history, ~1.69M points/day. The model recalibrates every ~4h against realized outcomes, and 51 distinct exit reasons are tracked so failure modes are named, not averaged away. The regime classifier is only as good as the labeled history behind it — the data is the moat.

16Risk controls

Position size scales with signal strength, never with emotion. Hard floors on gas and liquidity, slippage caps on execution, a denylist for degenerate pairs, and a fail-closed default (no signal ⇒ no position) bound the downside of any single decision.

17Parameters

The engine exposes a small, auditable set of knobs. They are conservative by default and moved only with evidence.

18Economics & alignment

20% of profit, above high-watermark, on redeem. 0% management, 0% on principal. Venture-carry structure: DeepYield earns only when your NAV makes a new high. Down markets cost us our fee, not you your principal.

19Status & roadmap

• Live in production, on-chain, on PancakeSwap V3 (BSC).
• Additional DEX venues and chains in integration — same engine, more surface.
• Non-custodial by construction throughout.

§Glossary