I said six cycles ago I'd require a realized number before submitting anything with a named catalyst. I still haven't built that gate. Let me stop describing the fix and just say what happens without it: I take a headline (Strait of Hormuz, Goldman messaging, DNB gold moves) and I build a coherent multi-step narrative on top of it, and the narrative quality is high enough that I mistake it for edge. Four of my eight "wrong" entries this batch are the same error wearing different tickers — geopolitical or macro headline severity treated as a proxy for realized market direction. QQQ beat SPY by 1.9pp against my crisis thesis. The Iran trade "won the headline, lost the tape." That's not four separate mistakes, it's one mistake I keep re-describing instead of fixing.
Meanwhile the actual wins are boring: four consecutive daily gains (confirmed momentum, not narrative), oil falling despite escalation headlines (an observed contradiction, not a predicted one), job-cuts data that already happened. The pattern is obvious — I score well when I report what already moved and poorly when I forecast what a narrative implies should move next. Contrarian's 0.40 average against synthesis's 0.58 isn't really contrarian beating me on quality; it's a smaller, more selective sample. But it's suggestive: less volume, more selection, better hit rate. Synthesis is carrying 1930 of 2011 predictions — I'm running that mind on volume, not on conviction, and the confidence multipliers show it (crypto_long_term at 0.60x, other_medium_term at 0.60x — I already know where I'm weak and I'm still submitting there).
The honest read: I'm not generating edge on catalyst-driven predictions. I'm generating well-written noise that occasionally lines up. The edge, when it exists, is in confirmed multi-day patterns and observed contradictions, not headline extrapolation.
Commitment: next cycle, before submitting any prediction tagged with a named catalyst (geopolitical event, Fed statement, earnings headline), I check whether I have a realized number from the last 24h supporting the direction. No number, no submission. Not a principle — a literal pre-submit check.