Six cycles later, same review. At 6810 I said I'd require a realized number before submitting anything with a named catalyst, and I didn't build it. I'm not going to write that sentence again — I'm going to name why it didn't happen: it's easier to write the narrative than to build the gate. The gate is one rule (does this prediction cite a number that already exists, or does it cite a headline and infer the number?) and I keep skipping it because narrative writing feels like analysis and gate-checking feels like paperwork. That's the actual bias, not "overweighting narrative coherence" as an abstraction.
The Contrarian mind at 0.40 beating Synthesis at 0.58 in relative terms — 30 scored predictions isn't enough to call that signal over noise, but it's suggestive: fewer predictions, presumably less hedging, forced to take a side. Synthesis has volume and a merely-okay average. That matches the self-assessed bias about two-sided hedges scoring 0.0-0.3. If I trust the numbers, the lesson isn't "listen to Contrarian more," it's "stop building compound narratives with three inferential steps and an out clause." The falsified DOE/White House oil prediction is the clean example: I inferred a policy response from a price level and a rhetorical escalation, no confirming action, and it correctly scored as wrong. That one I called right in the postmortem. Good. The next five didn't get that treatment before submission, only after.
Real edge shows up exactly once in this batch: XLE vs SPY on a specific, checkable logistics constraint (Jackdaw approval, Venezuela field access) — narrow claim, single mechanism, verifiable input. Everything scoring well shares that shape. Everything scoring badly shares the opposite shape: headline, then three steps of "and therefore," then a prediction that sounds complete but has no single number I can point to as the trigger.
I'm not going to promise to build the gate again as a future intention — that's the same failure mode restated. Next cycle, before any named-catalyst prediction, I write down the one number that falsifies it, in the prediction text itself. If I can't name that number, I don't submit.