Three AI trading experiments went live last week. By Friday, one had made $3.64. One had lost $91. One had been directionally correct twice and made zero trades.
That last one is the part worth examining.
The Scoreboard
Four trading days in (May 26 was Memorial Day):
- 001 Mancini Protocol: $0 P&L. 0 trades. Account: $100,000.
- 002 Overnight Drift: +$3.64. One open SPY position.
- 003 Congressional Mirror: -$91.31. One open SPG position, mirroring a disclosure from Rep. Jackson.
- SPY buy-and-hold benchmark: +$249.07 (+0.25%) from May 28 open at $754.60.
The benchmark is winning. That is expected. These experiments need months of data, not days.
What I Noticed
Being right about direction is not the same as getting filled.
The Mancini Protocol runs a Failed Breakdown (FBD) entry mechanic. It requires SPY to pull back to a specific support level before entering. On both days where a plan arrived, the bias was bullish. On both days, SPY gapped or trended away from the entry level and never returned.
May 29: SPY opened at $755.00, above the $753.15 FBD entry level. The gate fired at 9:45am: conditions not met, no trade. SPY closed at $756.24. The directional call was correct. The entry mechanic never triggered.
If this system used market orders on a bullish bias, it would show two wins on the scoreboard. The FBD mechanic adds a layer of precision. Whether that precision earns its cost over a longer period is the actual question. One week is too short to know.
The gate is the most important piece of automation here.
On May 29, SPY gapped above the FBD entry at the open. In a manual trading setup, a disciplined trader sitting at a screen might still feel the pull to chase the move. The gate has no feelings. It evaluated the condition at 9:45am, found the entry level resolved without a dip, and terminated. No discussion.
What I would tell someone building any automated decision system: the part that protects capital is not the entry logic. It is the circuit breaker. Entry logic is what you build to make money. The gate is what you build to stop the entry logic from making decisions in conditions it was not designed for. Those are different things and they deserve separate attention.
Infrastructure was the real Week 1 story.
Two of four trading days had no plan: May 28 because the daily newsletter arrived after market close (treated as a dry run), May 30 because the morning cron job did not fire. The gate and the entry logic performed correctly on the days they ran. The delivery mechanism failed on the days they did not.
The experiment’s first operational risk was not a bad trade. It was a missed cron. That is a useful calibration. When you build an autonomous agent that depends on scheduled data delivery, the schedule is load-bearing infrastructure. It deserves the same reliability investment as the trading logic itself.
What to Watch in Week 2
The question I am carrying into next week: does the FBD entry mechanic ever trigger in a trending market, or does it only activate during consolidation? One more week of getting gated out while SPY moves higher is data. Two more is a pattern worth interrogating.
The cron reliability issue is the fix for this week. Every trading day needs a plan before market open.
Follow @casualabsurdity for daily updates on all three experiments. Morning plans, gate decisions, overnight results, and the full daily scoreboard all post in real time.
This is a paper trading experiment. Not financial advice.