Demo — not yours · wallet-scoped numbers are a demo wallet · live market data stays live

Defend

Design a better risk curator.

Submit a program. Each step it reads market state — reserves, oracle price, utilization, clearable depth, flow, position distribution — and outputs the whole risk vector: liquidation LTV, borrow gap, collateral caps, delay length, oracle tolerance. We score it over 1,000 randomized simulations against hidden regimes it cannot see, and rank it by average edge. Win on-chain and your policy becomes protocol infrastructure — public, forkable, your name on it. Run a private model offchain and charge for it: compute costs are the only real fee moat, because anything fully on-chain can be copied the day it wins. The house entries below are real venue policies; the challenge is beating Membrane’s own algorithm.

starter repo ⧉
read a curator→allocate across curators→become one
01 /

The decision surface

every lever your program outputs · move any of them and watch the couplings
Liquidation LTV (M)86%
Borrow gap (g)3%
Collateral cap$30M
Delay window
Oracle staleness
Borrow LTV — B = M − g83%
Wipeout frontier √M92.74%
Headroom, trigger → wipeout3.3 pts
Capital efficiency vs Aave BTC+11% borrowable per $
Max safe delay at this M — crash depth vs headroom>48h
Your delay8h — inside the safe bound
House cap formula — depth × (1 + junior − vol stress)$29M
teal tick = formula cap · your cap past it is cascade exposure$1M cannot be cleared in a cascade
Oracle: max price error at liquidation±0.90%
Oracle: expected frozen time per week2 min — tolerable

Every number is a consequence, not a setting — and the levers couple. A delay that is safe at 86% is lethal at 95%. A cap that ignores clearable depth only fails in the cascade regime. Your program outputs this whole vector, per collateral, every step.

02 /

Leaderboard

avg edge over 1,000 sims · house policies pinnedi

1

Depth-Led Adaptive v2

0x7a41…9c2e

+3.86

14 attempts

2

B = M − g (house controller)

membrane · house policy

+3.41

—

3

Vol-Gated Caps

0x22b0…1f77

+3.12

31 attempts

4

submission #3

you

+2.48

3 attempts

5

86% fixed

morpho (static) · house policy

+2.21

—

6

78% BTC / 83% ETH fixed

aave (static) · house policy

+1.74

—

7

90.91%, ETH only

liquity-style · house policy

+1.32

—

regime strip per row: edge by hidden regime — calm · vol spike · depth shock · cascade. Red = negative in that regime. mock

03 /

Where your last submission bled

per-regime breakdown — this is the practice loop, not a verdict
Calm drift · 412 sims

held with the pack

+3.1

Vol spike · 305 sims

your caps tightened in time

+3.3

Depth shock · 178 sims

edge thinned — caps lagged the depth draw

+2.2

Cascade · 105 sims

bad debt: LTV held high into forced flow. This regime alone cost you rank 2.

-1.6

submission #3 · scored 2026-08-15 over 1,000 draws · seed set rotates per scoring epoch

04 /

Run one sim before you submit

05 /

Mirror vaults — two classes, one honest split

on-chain policy is copyable, so it cannot carry a fee · offchain compute can

Depth-Led Adaptive v2

frozen
Author0x7a41…9c2e
Sim edge, rank 1 of epoch 4 mock+3.86
Deposits$412k · 37 depositors
Classpublic policy — fully on-chain
Feenone — copyable code cannot charge rent
Author earnsrank, provenance, and the listing it seeds
Updatesnone — frozen means the program, not the params

A frozen program still adapts — it was scored on reacting to regimes it had never seen. Fork it freely; that is the point of winning on-chain.

Operator vaults

offchain · fee-bearing
Why a fee holdsthe model runs offchain — it costs money and cannot be read out
Disciplineevery update re-simulates before it activates
Depositor viewsame divergence chart, same rules

Live performance against the sim expectation band. The divergence is shown because it exists: sim edge is not live edge, and the drift between them is the honest number an allocator needs. Band = middle 80% of the 1,000 scored sims.

06 /

What the sim cannot see

●Venue recalls are not modeled. Live positions can pull deployed capital back to repay debt before collateral sells. The sim has no such lever, so live outcomes should beat sim outcomes in shallow breaches. If they do not, that is signal.
●Sim edge is not live edge. Calibration drifts from the moment it is measured. The vault chart above is the correction, updated as live data settles.
●Rebreach timers are simulated both ways — with and without reset — because the shipping behavior is an open decision. Your score shows both columns.

calibration: backtest v5 · 27 files · Chainlink rounds · 3,111 Aave + 886 Morpho liquidation events · protocol params at block 23,543,615 · USDe trade-level data · Curve depth

07 /

Submit

starter repo: github.com/membrane-fi/curator-starter ⧉

sandboxed runtime · scored on rotating seed sets · submission cooldown applies

08 /

Decisions pending — not resolved here

1DIRECTION SET — two tiers: on-chain winners are public policy (free, forkable, seed listings); offchain operator vaults carry the fees, because compute cost and unreadable models are the only durable moat. Remaining open: the exact split of what each tier may control.
2Fee shape for operator vaults (on-chain tier now feeless): ongoing vs performance — still open.
3Bad-debt penalty coefficient in the edge formula.
4Runtime: WASM preferred; languages supported.
5Managed vaults at v1? Each update invalidates the sim record; if offered, re-simulation before activation is mandatory.
6Rebreach timer reset — both variants simulated; one ships.
7Anti-gaming: rotating seed epochs and cooldowns proposed; not locked.
the carry board → /carry