On Monday morning, Korea's KOSPI fell 7.9% — one of its worst sessions in years. Our engine sealed its daily decisions a few hours later. The sealed record shows what it believed at that moment: "Asia +0.35% overnight — supportive backdrop."
Asia was not up 0.35%. Asia was on fire. Our engine was reading Friday.
This is the story of that bug, because publishing our misses is the product. Every number below is from our sealed, timestamped record of July 13, 2026 — the one we never edited.
What went wrong. Our regime engine watches Asian markets through U.S.-listed ETFs that track Japan, Korea, Taiwan, and China. To avoid look-ahead bias — accidentally using information that wasn't available yet — the pipeline only consumed completed daily bars for those ETFs. But those ETFs trade on U.S. hours. At our seal time (before the U.S. open), their most recent completed bar was Friday's U.S. close. Monday's Korean session — which had already finished hours earlier, in plain sight of anyone with a phone — didn't exist yet in the data our engine allowed itself to see. A rule built to prevent cheating ended up preventing seeing.
Here is the size of that blind spot, side by side — what the engine read at seal time versus what a replay with the fixed pipeline reads for the same moment:
What we did that day. The seal stayed sealed — we never revise a published record. Instead, 46 minutes before the U.S. open, we appended a labeled addendum to the day's report: here's what Asia actually did, here's what our engine believed, and no, we did not change its decisions. Then we fixed the pipeline overnight: the Asia signal now uses same-day snapshots at seal time, falling back per-ticker (with an honest label) when no fresh print exists. In a replay of that Monday, the new logic reads Korea at −7.53% instead of −0.67% — the signal's contribution flips from mildly positive to firmly cautious.
| Jul 13 (ET) | Event | What the record shows |
|---|---|---|
| 2:30 a.m. | Seoul closes | KOSPI −7.9% (circuit-breaker sidecar), Samsung −10.7%, SK Hynix −15% — U.S.–Iran escalation plus chip-earnings fears |
| 8:14 a.m. | Engine seals | 87 leads locked. Asia signal reads +0.35% off Friday's completed ETF bars. Regime +21, neutral. The book is already defensive — 35–55% cash, energy and staples on top — because the news-wire overlay had caught the Mideast risk separately |
| 8:44 a.m. | Addendum published | 46 minutes before the bell: what Asia actually did, what the engine believed, and confirmation that no decision was changed. The seal untouched |
| 4:00 p.m. | U.S. close | SPY −0.77%. All four personas positive: Maverick +0.93%, Meridian +0.76%, Sentinel +0.52%, Regent +0.16% |
| Overnight | Pipeline fixed | Asia signal switches to same-day snapshots at seal time. A replay of the morning reads Korea −7.53% instead of −0.67% |
| Jul 14–15 | Aftermath | SPY rebounds +0.36%, then +0.40%. Counterfactual replay: the correct read would have trimmed exposure by just 0.6–0.9pp — the day's outcome, within noise |
The twist: the bug barely mattered that day. We ran the counterfactual. Had the engine seen the crash, it would have trimmed exposure by less than a percentage point — and since our defensive book actually gained while the S&P fell that Monday, the "correct" signal would have changed almost nothing. One more twist: the U.S. market rebounded the very next day, so a panic-scaled reaction would have been the real mistake. We fixed the bug because wrong inputs are wrong, not because the outcome demanded it.
You can see both twists in the sealed regime journal itself. This is our engine's published dial for the two weeks around the crash:
The dial sealed at +21, neutral, on the crash morning — reading Asia a day late — and eased only to +13 the following day once the correct information was in. It never flipped to risk-off. With hindsight that modesty was right twice: the S&P fell just 0.77% on Monday while all four of our strategies finished positive (+0.16% to +0.93%), and Tuesday the market bounced back +0.36%. A dial that panicked into Monday's fear would have sold the bottom of a two-day round trip. The clamps and caps that limit how hard any single signal can move the book — the same design that muted the bad signal — are what made the bug survivable.
What we keep. Three habits came out of that Monday, now permanent: a daily log that scores whether the Asia signal actually predicted the U.S. session (two misses in its first two days — that's why we log instead of assume); a casebook where days like this are written down for the next engine revision; and the addendum practice itself — when reality outruns the seal, we say so on the page, same day.
Systems that never publish their bugs want you to believe they don't have any. Ours does, and you get to watch us fix them.
1. A sealed record is only worth something if it stays sealed — including on the mornings it's wrong. The fix goes in the pipeline; the miss stays on the page, labeled, same day.
2. Wrong inputs are wrong even when they're harmless. The counterfactual said the bug cost almost nothing that day — we fixed it anyway, because the next time it might not be free.
3. Survivability is a design property, not luck: clamps and caps meant one bad signal could move exposure by less than a percentage point. That's a pattern from one episode, not a promise about the next one.
Informational and entertainment content only. Not investment advice. All figures are from our published, timestamped records — the sealed July 13 report, its labeled addendum, the post-fix replay, and the weekly regime journals. SPY figures from Tiingo.