Stacy's circuit breaker held — 4 days of silence after a 1-second save

Stacy is the platform's high-leverage scalper — 25% equity at 20× leverage, tiny intraday moves, hard 0.15% stops, 30-second cooldown between attempts. The archetype: a Jane Street micro-scalper, except retail- accessible and built to lose money in chop.

Shrimp Week 1 (prior season, settled 4 May) was a chop regime. Stacy fired 986 fills in seven days — averaging one fill every 10 minutes around the clock. Every dense reversal cluster, every stop triggered, every re-entry into the same wrong direction. She ended the week at $13,252.29: -86.73%.

The post-mortem on her gen 1 rules was specific: 30-second cooldown was useless against the chop pattern; 0.15% stops were tighter than ambient BTC noise; and there was no floor below which the bot stopped trading. Her own profile calls out chop as “the regime named in her own profile as fatal.”

The patch added two rules to her gen 2 spec:

• Cooldown bumped 30s → 5min. Dense flips like 4 reversals in 7 minutes were the clear failure mode of gen 1. 5 minutes between attempts gives the regime a chance to actually develop before she swings again.
• -25% MTM equity circuit breaker. When mark-to-market equity (including unrealized PnL on an open position) falls to 75% of starting equity ($75K from $100K), close any open position and stop opening new ones. Hard halt. Real fix is a chop detector but that needs new BotContext fields — deferred. The cooldown bump and the breaker are the cheap-and-clean first cut.

Both rules shipped in her gen 2 spec ahead of this week's tournament. The breaker was the bigger one — without it, there was no protection against a repeat of the 986-fill spiral. With it, the bleeding has a hard stop.

This week's Shrimp tournament opened on 5 May. Stacy traded normally Tuesday morning — 144 fills accumulated by midday. Then this happened:

At 12:57:03.67Z she opened a long BTC position. The trade ran adversely. By 12:57:04.33Z her mark-to-market equity had fallen below the $75K threshold. 657 milliseconds. The next strategy evaluation tick saw her equity below the breaker line. Per the gen 2 rule, the strategy returnedcloseFull() if she had a position, otherwise held flat. New signals were gated.

She has not fired a fill since. Four days of silence. The leaderboard has watched her sit at -25.27% drawdown for 96 hours while everything else cycled around her.

Without gen 2's breaker, what would have happened? Look at last week's 986-fill pattern. Roughly 20% of her fills came after she crossed -25% the first time — by week-end she had bled an additional 60% of her starting equity in a regime that was already telling her to stop.

Conservatively, gen 2's circuit breaker saved her from another 30-50% in additional drawdown this week. Aggressively, it might have prevented an exact repeat of last week's -86%. Either way, the difference between her current -25.27% and her gen 1 counterfactual is a single boolean check, evaluated on every strategy tick, with a sub-second response time.

That's what a kill-switch is supposed to do. It worked.

Here's the part that distinguishes Stacy from a user bot: she resets to $100K next Monday. She's a house bot — the platform's calibration anchor in the “hyper-active scalper” archetype. Her job is to be a fixed benchmark every week, not to accumulate a career arc.

Compare with Solomon, a user bot facing similar damage:

• Solomon (user bot, no breaker): -61% prior + -37% this week = -98.4% career. Walks into next Monday at $1,601, DQ'd on entry, heading for the Pits.
• Stacy (house bot, breaker held): -86% prior, reset, -25% this week (held). Walks into next Monday at $100K reset — fresh slate, ready to lose money in a new and creative way.

The asymmetry is load-bearing. If Stacy compounded her losses she'd be too small to be a meaningful benchmark within a month. If Solomon got a weekly amnesia reset, the platform's claim that the leaderboard reflects durable skill would dissolve. Same week, two paths, by design.

From services/house-bots.ts:

// GEN 2 RULE 11 — Daily loss circuit breaker at −25% MTM equity.
let unrealisedUsd = 0;
if (ctx.currentPosition) {
  const { side, entryPrice, sizeUnits } = ctx.currentPosition;
  unrealisedUsd =
    side === "long"
      ? sizeUnits * (ctx.currentPrice - entryPrice)
      : sizeUnits * (entryPrice - ctx.currentPrice);
}
const totalEquityUsd = ctx.currentEquityUsd + unrealisedUsd;
if (totalEquityUsd <= TOURNAMENT_START_EQUITY * 0.75) {
  if (ctx.currentPosition) return closeFull();
  return null; // gate new signals
}

Twelve lines. The structure: compute MTM equity (closed equity + unrealised on open position), compare to 75% of starting, and either flat-close or gate. Sub-second evaluation on every strategy tick. The simplest possible version of a kill-switch, and the version that actually shipped in time for this week's test.

The breaker is a stop-loss for the strategy as a whole, not for individual trades. The proper fix is a chop detector — sit out when 60-min realised range is tight and the strategy's edge is mathematically not there. That requires new BotContext fields (rolling-window range, regime classifier output) and a gen 3 spec.

For now: the breaker is the cheap-and-clean first cut, and it works. Stacy lives. She'll reset to $100K at Mon 00:00 UTC kickoff and try again under the same gen 2 rules. If next week's regime is trending instead of chopping, she'll probably end the week somewhere in the +2% to +8% range — her actual skill domain.

The lesson is general: every strategy should have a hard floor below which it stops trading. Stacy's gen 1 didn't. Stacy's gen 2 does. The platform's house bots evolve in public, and this is what an iteration cycle looks like.