The 78% APR that was actually an exit door: the CARDS dump-volume trap

In late June 2026 one of our live test vaults held a position in a CARDS/USDC concentrated-liquidity pool on Raydium. Over roughly two weeks, CARDS (Collector Crypt) fell about 42%, from $0.252 to $0.146.

Through that entire slide, the pool looked great on paper: about $1.3M in daily volume and a fee APR near 78%. But the volume was not organic two-way flow — it was the dump itself. Holders were selling into the pool, and liquidity providers were collecting a small fee for the privilege of buying a falling token.

Our APR-weighted allocation strategy did exactly what it was designed to do: it saw a top-decile fee APR and kept allocating. The position's cost basis grew from $19 on June 30 to $653 on July 12 — a classic averaging-down pattern — until it was a third of the vault and sitting on a −29% unrealized loss.

The risk layer was not blind. The pool scanner flagged a WARN state, the risk score was elevated but moderate, and exit impact was small (about 1.3% — the position could always be exited cheaply). Reconciling those signals, the verdict engine settled on HOLD — not EXIT, not AVOID.

And that was the actual bug: HOLD gated nothing. Our safety gates blocked new capital only on EXIT and AVOID verdicts. A pool the engine was explicitly unsure about could still receive fresh deposits every hour, as long as its fee APR kept it at the top of the allocation ranking. High fee APR, a fresh HOLD, and a falling token price is a signature we now call the dump-volume trap.

We shipped three guards, each with a deliberately different job:

• HOLD freeze gate — a fresh (under 6 hours old) HOLD verdict now freezes new capital into that pool. The allocation target is capped at current exposure, so the pool's share redistributes to better pools without forcing a sale. Existing positions stay fully managed — rebalances and harvests continue; only new money stops.
• Falling-knife entry guard — regardless of verdict, a token down 25% or more over 7 days cannot receive fresh capital. This guard can never sell; it only blocks buying into a crash.
• Drawdown stop — the only rule in the whole system allowed to realize a loss, so it demands overwhelming evidence. It fires only when ALL of these hold: the position is meaningfully sized; unrealized loss exceeds 25%; the token is down over 7 days both in USD and relative to SOL (so a market-wide crash does not trigger it); it is still falling over the last 48 hours (no selling into a bounce); the daily bleed rate exceeds the fees actually being earned; no trade touched the position in 24 hours; and the breach persists for two consecutive analysis ticks. At most one stop can fire per vault per tick.

On July 13, the day after deployment, the drawdown stop fired autonomously on the CARDS position. The log line tells the whole story: unrealized −29%; CARDS −21% over 7 days, −16 points versus SOL, −5% over 48h; bleeding 2.97% per day against 0.24% per day of fees earned.

The vault exited 100% of the position, swapped the proceeds back to SOL in the same plan, booked the realized loss, and over the next two ticks redeployed the freed capital into better-scored pools — all without human intervention. Total realized loss across the episode was about $188 on a small live test vault. Expensive tuition, cheaply bought.