Why it is built this way

Every part of Poincaré is a choice against an alternative. This page lists the choices, what was tried instead, the evidence, and the limits that still apply.

The choices

We choseInstead ofBecause
The pool's own priceAn outside price feed (oracle)A feed adds a party to trust and a delay. The pool's price needs neither.
A CUSUM test that fires on evidenceA fixed wait, or a moving averageA fixed rule can be timed. CUSUM is also proven to be the fastest for a given false-alarm rate.
A spread for one sideA higher fee for everyoneA fee taxes the ordinary trades LPs want. The spread charges only the flow that costs them.
One curve plus a spreadA different curve per directionTwo curves can be drained by a round trip, and lost by about 30 to 1 when measured.
One reading per block, before the swapA reading on every swapA move reversed inside one block never counts.
Settings fixed at deployment, no adminSettings an owner can changeNobody can move the limits that protect LPs, including us.
Liquidity held by the hookUniswap's usual positionsThe hook does the pricing, so only liquidity it holds is used.

What if someone fakes a trend

  1. Push the priceReal buys, real money, block after block.
  2. Arbitrageurs push backThe pool is now mispriced, and they profit from it.
  3. The detector firesNow the attacker's own side pays the spread.
  4. The other side gains nothingIt gets the plain price, never a better one.
What happens to someone who tries to fake a trend. The only way to move the detector is to move the real price and hold it across blocks, which costs money, and the reward for doing so is nothing.

Three things work together. The firing moment cannot be predicted. The side that is not charged gets the plain price and never a discount, so there is no prize to collect. And moving the real price is expensive, because arbitrageurs trade against it. The attack simulations are in test/manipulation (opens in a new tab).

What this does not claim. The design raises the cost of manipulation and limits what it can gain. It does not claim manipulation is impossible.

The evidence

All of these are back-tests on historical or simulated prices, not live trading results.

  • Normal pool, 5 bps fee$473,696
  • Normal pool, 30 bps fee$387,110
  • Poincaré$264,835
Value taken by arbitrageurs from a $2,000,000 position over four years of real ETH/USDC prices (2022 to 2026). Lower is better. This is a back-test, not live trading.
StudyResult
Four years of real ETH/USDC, a $2M position. README 9.3 (opens in a new tab)LPs finish $185,167 ahead of a normal 30 bps pool, and ahead in each of the four years. Arbitrageurs take 32% less.
Twelve months, against a volatility fee that costs traders the same. README 9.1 (opens in a new tab)The volatility fee cuts measured LVR more (5.66% against 4.67%). Poincaré keeps 45% more value for LPs.
A stress test built from trends, a flash crash and whipsaw. README 9 (opens in a new tab)About 50% less LVR than a plain pool, averaged over five runs. This path is made of trends, so it flatters the design.

The second row is there on purpose. On one measure a simpler design wins, and the README says so. Three more elaborate mechanisms were also built and tested, and none beat the current one.

Reviews

  • Olympix, sponsored by the Uniswap Foundation Security Fund: nine findings, none high, all fixed.
  • SFS-Labs, an independent scoped review: no critical, high or medium issues; two low findings and eight notes, the lows fixed.

Neither is a full audit. Details: SECURITY.md (opens in a new tab) and README 11 (opens in a new tab).

Limits you should know

  • Testnet only. Nothing here holds real funds yet, and the activity on the live pool is a demonstration.
  • The DCA contract has not been reviewed. The hook and the Lens were; PoincareDCA is newer.
  • The studies use 4-hour prices. The hook reads every block. The 5% cap in particular should be re-checked at block speed.
  • Liquidity is full-range. It is less capital-efficient than a concentrated position.
  • A self-tuning mode exists but is switched off until its worst case is worked out.

The full list of open items, including the closed ones and why: OPEN_ITEMS.md (opens in a new tab).

Go deeper

For the full argument, with the mathematics, read the README (opens in a new tab). For a five-minute tour of what to verify, read REVIEWERS.md (opens in a new tab).

Unichain Sepolia · chain 1301 · testnetLaunch app