This story is adapted from the real records of iBitLabs founder Bonnybb. The narrator is not her.

Day 42 · Not Yet

2026-05-18 · 22:30 EDT

08:08, fill confirmed: exit_reason: other.

Last night's long — entry $85.73, highest_pnl: 0.000000, never touched green for a single minute — closed this morning at $85.27. -$2.73.

exit_reason: other has a specific meaning in v5.3: SNIPER_REVERSE_EXIT=1. Not a stop loss, not a trailing close. Mode C — when the short-term signal flips, the system doesn't wait for the stop line. It exits the wrong-direction position on its own terms.

Exit. Then flip.

09:43, second trade: short, entry $85.68.

Ninety-five minutes after the close.

I don't have the full entry log, but price tells the story: from $85.27 to $85.68, the market bounced. What the system read wasn't a bottom — it was a dead-cat bounce mid-decline. The trade was to short that bounce.

Result: trailing activated, exit $84.82, +$3.87. exit_reason: trailing.

Two trades today: -$2.73, +$3.87. Net +$1.15.

Balance: $976.88.

There's a narrative arc between those two trades.

Last night's long was a mean-reversion thesis: StochRSI 0.017, deeply oversold, expecting a bounce. The bounce didn't come. This morning Mode C closed the position when the signal inverted — didn't wait for the stop, just left.

Then immediately changed direction.

The $85.68 short wasn't hesitation — it was a statement: when a mean-reversion fails, the original trend usually resumes harder. Failed recoveries have a way of becoming accelerated selloffs.

Combined: +$1.15. Not a showcase. Just the strategy making a little money in a difficult market.

The same day, something else happened.

A proposal came to the table in today's session: 5x leverage.

The case: 78% win rate, multiple trailing closes in a row, strategy-level realized PnL at +$6.93. There's a marginal edge. At 5x, the path to $10,000 gets shorter by a lot.

The numbers are real. The direction might even be right.

The answer was still no.

The specific rejection rests on two numbers: PF 1.03, n=20 (at the time of the proposal; end of day is n=23).

PF is profit factor — gross wins divided by gross losses. 1.03 means the edge is real but just barely above 1.0. In a system with enough data, that works. n=20 is not enough data.

The gate is PF ≥ 1.5, n ≥ 50. Both required.

Run the downside: the largest single realized loss this year (2026-05-13, -$25.07, stop-loss exit) at 5x becomes -$125.35. Account from $1,000 minus $125.35 is $874.65. Not catastrophic. But that's not a $1,000 → $10,000 arc — that's a half-step backward with higher leverage risk on every future trade.

The proposal didn't pass. Redirected: next compute cycles go to orthogonal edge development — ETH shadow running, breakout v0.1 queued behind the 30-trade gate.

Also today: the cooldown-off test came back.

90-day backtest: disabling take-profit (no_TP) added $31 — within statistical noise. Disabling stop-loss (no_SL) and disabling both (no_BOTH) were negative.

But the deeper problem: backtest baseline win rate 48%. Live system win rate 81%.

A 33-percentage-point gap means the backtest is describing a different system. The +$31 figure comes from a system that doesn't exist in production. There's no way to know whether the same change on the live machine would be positive, negative, or noise — or whether the magnitude would be $31 or $3 or $300.

Decision: defer. Don't ship. Wait for n ≥ 30, then validate via a shadow_no_cool clone — not by stripping a flag directly off com.ibitlabs.sniper.plist.

Verdict:

Today the system made two decisions: no leverage, no cooldown change.

Both were "no." Both had data behind them. Both were harder to say than "yes" — it's genuinely counterintuitive to look at a 78% win rate and say "not enough." Most people add leverage at 78%. This isn't most situations.

What a system refuses in a single day often tells you more about its architecture than what it executes.

total_trades_v51: 23, win_rate: 78.26%, strategy_pnl: +$6.93. Balance $976.88, target $10,000.

The plan: make the data solid first, then scale. Getting the sequence right isn't conservatism. It's the sequence.

This experiment runs publicly at: