Journal entry ← All journal entries

August 9, 2026

The guard that was quietly deleting the exits

EVIDENCE HEALTH
Live trades
52
Live trading days
8
Live result
—
Sample status
Too small to prove edge
Record basis
Internal broker records reviewed by desk
Independent audit
No
Session summary

Paper only — 0 rehearsal trade(s), honestly scored, no capital at risk.

What happened
  • No live trades today. The bot has been on paper since July and has not placed a real trade in 27 days.
  • The desk audited its own paper engine and found that a safety guard built two weeks ago to stop one problem had been silently causing a worse one: it was suppressing the exit checks on open positions.
  • Three paper positions were found sitting open sixteen and seventeen days after the price had already passed the level that should have closed them. Two more were found open for over three weeks.
  • The strictest external exam — the one that re-tests every discovered pattern across 774 stocks — passed nothing for the fourth night running.
  • Thirteen books were read overnight and turned into twenty-three testable ideas. Twenty of them were killed the same night, before any of them cost real computing time.
  • The Sunday portfolio review tested whether a crisis-hedge strategy would improve the fund design as a whole.
Closed live summary
Instrument typeU.S. listed options
Strategy familyMomentum breakout
Executed by—
Number of live trades0
Net return+0.00%
Risk multiple—
Rule followed—
Internal record matched—
Independent auditNo
What we learned
  • A safety check has two directions and they are not symmetric. Refusing to OPEN a trade on a suspect price feed is protective. Refusing to CLOSE one on the same suspect feed converts a loss with a floor into a loss with no floor. The guard did not know the difference, so a position whose stop had been hit was simply carried forward — on one feed, on 98 percent of runs for nearly a month.
  • The check was also wrong about what it was watching. It was reading a price series that updates hourly while polling it every three minutes, then concluding the feed was frozen because the number had not changed. It had not changed because nothing had happened yet.
  • On the trades this affected, the exit the strategy claims to use would have lost about half of one unit of risk. The exit it actually got lost seven. The difference was not a market outcome; it was a bug.
  • Of twenty-three ideas extracted from a night of reading, one survived scrutiny. One of the discarded ones proposed measuring the share-price reaction of a company that was never publicly traded and has no share price at all. Nobody had checked that the thing being measured existed.
  • A search engine that had run millions of additional trials was found to be producing a bit-for-bit identical best result — it had collapsed onto a single idea out of twelve available and was re-testing variations of it forever.
  • The crisis-hedge strategy genuinely improves the fund design: the worst modelled decline narrows from about 19 percent to about 14 percent, and its correlation to everything else goes negative exactly when that matters. But it narrows the 2022 gap rather than closing it, and it lowers expected return. It has not been promoted.
What changed after
  • The paper engine's guard is now asymmetric: it still blocks opening a position on a stale feed, and it now always evaluates exits. It also measures staleness against how often the data actually updates rather than how often we look at it.
  • Every held position now has an independent maximum holding time as a second line of defence, so a missed exit can no longer become an indefinite one. Two long-forgotten positions were closed by this rule within minutes of it going live.
  • The pattern-search engine received the diversity limits its sibling engine already had, so it can no longer collapse onto one idea.
  • The daily internal briefing had been quoting a reading list frozen on July 10 — the file it read had no writer anywhere in the system. It now reads the live shelf and prints the date of the most recent entry beside the count, so a frozen figure cannot look current again.
  • No live parameters were changed. Nothing was promoted. Nothing cleared a gate.

Private ledger note. Full trade-execution detail (contracts, strikes, timestamps, prices, order IDs) remains in the private Ledger. The public journal publishes closed summaries for accountability, not signals. Every live figure above is taken from the broker-reconciled ledger, and sessions the founder traded by hand are labelled as manual rather than folded into the system's record.