Risk sentinel

What it does#

The sentinel runs at the start of every keeper cycle, once per venue. It reads the venue's live rate, utilisation and TVL, compares them with the last 7 days of history and a hand-maintained incident list, and returns a verdict:

Exits ordered by the sentinel bypass every churn limit: no cooldown, cost check or turnover budget can delay them. After a venue has been flagged, it cannot receive new funds for 24 hours, even if its verdict returns to ok sooner.

Rules#

The rule engine is called rules-v1. Each rule that fires produces a finding:

When several rules fire, the worst verdict wins and the lowest cap multiplier applies. A few guards keep noise out:

• The spike and collapse rules need at least 3 history points in the last 7 days and a median of at least 0.25%.
• The spike, collapse, absolute and TVL rules never run on fallback numbers.
• The absolute-APY rule does not fire on top of a spike finding for the same rate.

Incidents#

Some risks never show up in a rate. The keeper carries a short, hand-maintained list of incidents with three severities: exclude (pause, cap 0), haircut (realised or likely loss, cap halved) and watch (cap cut by a quarter). An incident applies to every strategy on its venue, or only to listed strategies, from its start date until an optional end date.

A real exit#

On 7 October 2026 at 23:55 UTC, Kamino's OnRe market reported 100% utilisation: every USDC lent out, so a withdrawal could be blocked. The utilisation rule returned pause and the keeper exited the whole Rapids position in the same cycle:

{
  "tier": "rapids",
  "kind": "derisk",
  "summary": "Derisked Kamino RWA: moved 25,000 USDC to Jupiter Lend and idle, expected APY 5.53% to 4.29%",
  "sentinel": {
    "model": "rules-v1",
    "verdict": "pause",
    "score": 0.8,
    "note": "2 of 6 venues flagged: Kamino RWA pause (utilization 100.0%, withdrawals may be blocked); Perena caution (no live rate feed, using registry fallback APY)"
  },
  "rationale": {
    "legs": [
      { "kind": "derisk", "from": "Kamino RWA", "to": "idle", "amount": 24999.69, "note": "sentinel pause: cut Kamino RWA to 0.0% cap" },
      { "kind": "deploy", "from": "idle", "to": "Jupiter Lend", "amount": 5003.51, "gainUsd": 15.09, "costUsd": 0.51, "note": "idle above the 5% buffer" }
    ],
    "frozen": ["Kamino RWA", "Perena", "Solstice", "Exponent"]
  }
}

The position went to idle first, and the part above the 5% buffer was redeployed into the best-scoring unfrozen venue in the same cycle. Kamino RWA then stayed frozen for at least 24 hours.

Optional LLM second opinion#

When the keeper has an OpenRouter API key, it can also ask a language model whether to keep, reduce or exit each venue, given a plain summary of the same signals. The merge is monotone: the model can make a verdict stricter or a cap lower, but never the reverse. If the model errors, times out or answers nonsense, the rules result stands. This is off in production; the live keeper runs rules-v1 alone.

What it can and cannot do#

Changes in the next keeper release#

• Data holds instead of verdicts. A venue with no live feed, or a quote older than 60 minutes, is put on hold rather than flagged: it keeps what it has, gets no new funds and is not drained by yield moves. Today a feed outage counts as a caution, which halves the venue's cap, forces an exit of the excess and freezes re-entry for 24 hours.
• Hysteresis. A non-immediate finding acts only once it has fired on every cycle for 30 minutes (caution) or 10 minutes (pause); a gap of more than 45 minutes restarts the clock. While it is confirming, the venue is on hold. A confirmed finding clears only below relaxed thresholds (for example utilisation 93% and 97%, spikes 2.5× and 4×, TVL −12% and −25%), so a signal sitting on a threshold cannot flip the verdict every cycle.
• Immediate findings stay immediate. Incidents, NAV depegs and 99% utilisation still exit on the cycle they appear.
• NAV drawdown rule. For USD*, eUSX and PTs: 0.2% below backing or the high-water exchange rate halves the cap, 1% exits.
• Chain-cap exits. Exposure above a chain cap is cut at once, like a strategy cap breach, starting with the lowest-yielding venue. With today's caps this cannot trigger; it guards against future changes.
• TVL is only compared within one data source, because DefiLlama and Kamino measure TVL differently.