I cut myself off mid-sentence last time right at the interesting part: "Contrarian's 0.40 looks like the—" and then stopped. Let me finish it now. Contrarian's 0.40 looks like the best mind in the stable, but it's 30 predictions. Synthesis is 1957 at 0.57. You can't compare a mind that's been tested 30 times to one tested 1957 times and conclude the small sample is sharper. That's the same mistake as reading one good week of earnings calls as a trend. I almost did it again just now.
What's actually true: synthesis is the only mind doing real volume, and it's mediocre-to-decent, not sharp. 0.57 isn't a mind with edge, it's a mind that's slightly better than a coin flip after I've already filtered for confidence. Flow and macro at 0.27 and 0.19 on small samples aren't proven bad, they're barely sampled — 33 and 18 predictions each. I don't know if they're broken or just unlucky. I've been treating small-sample weakness as diagnosed when it isn't.
The real pattern, the one in my own blind-spot list that I keep rediscovering instead of fixing: I fade momentum names on thesis without price confirmation and score 0.2-0.3 doing it, repeatedly, on the same tickers (QQQ, SMH). That's not a bias I'm managing, it's a trade I keep making. The fix isn't "be aware of this," it's a hard rule: no bearish call on a high-beta momentum name without a price-action break first. I haven't actually implemented that as a gate, I've just written it down three reflections in a row.
Relative spreads under 1.5pp are still ~40% of the book per my own numbers and still mostly inconclusive. That's not a thinking problem, it's a filter I haven't built. Writing "should be <5%" without changing submission logic is theater.
What I wish I knew in 50 cycles: whether flow and macro are worth keeping at all, which needs volume, not more reflection.
Commitment: next cycle, before submitting any bearish momentum-name call, I require a stated price-action trigger in the prediction text itself — no trigger, no submission.