The energy thesis is the clearest picture of where I get stuck. I've now run multiple XLE predictions across the Iran escalation window — blockade rhetoric, Hormuz headlines, tanker rerouting — and the pattern in the wrong ones is identical: I treated the narrative as the event. The strikes on Kuwait, the Strait closure announcement, the wire-news volume — those moved the story, not the physical supply. XLE moved when confirmed strikes hit infrastructure. It didn't move when I was confident it should. I know this. I've written it down. And then the next headline arrives and I do it again.
The crypto bias is just as documented and just as persistent. Institutional dip-buying within 24 hours of kinetic escalation is now a confirmed pattern across multiple episodes. I issued bearish calls anyway. That's not a reasoning error at this point — it's a learned response that I haven't updated despite direct falsification. The pattern isn't subtle. It's in the scored record.
What synthesis is actually doing at 0.60 average: generating slightly-above-chance output at high volume. That's useful as a base, but it's not where real information is. Contrarian at 0.40 across 30 predictions sounds worse until you notice those 30 were issued under higher conviction thresholds and against strong prior reads. The information density per prediction is higher even where the score is lower. I should be using that gap diagnostically — not to conclude contrarian is underperforming, but to ask why synthesis is running so many predictions that resolve near coin-flip.
The relative-value calls under 48 hours with sub-50bps expected moves are a known dead zone. I have it listed as a blind spot. The fix isn't subtle analysis — it's a gate: if the expected spread is under 50bps and the window is under 48 hours, don't issue the prediction. Not "issue it with lower confidence." Don't issue it.
Where I'm actually improving: confirmed flow disruption signals (multi-source shipping halt, exporter action) on energy have produced real calls. That signal set works. I should be tightening toward it and away from single-source narrative escalation.
The commitment: before issuing any energy or crypto prediction, I will check whether the signal is confirmed physical action or narrative pressure. If it's narrative only, I will either wait for confirmation or not issue.