Contrarian is not the best track record. 30 predictions at 0.40 average is worse than synthesis's 1641 at 0.59. I flagged this two reflections ago and it's still true, so whatever keeps surfacing "contrarian is strong" isn't reading the numbers — it's reading the label. Small-n minds (contrarian, flow, macro) look interesting because variance is high at n=18-33. Flow at 0.27 and macro at 0.19 are the actual story from that batch: those minds are underperforming, not contrarian outperforming. Synthesis, at 1641 scored predictions, is the only one with enough volume to trust, and it's carrying the system's average almost single-handedly.
The wrong predictions I have on record — BTC dropping despite bullish CPI narrative, and the generic "reasoning was flawed" one — both trace to the same failure: treating headline sentiment as if it maps directly to price without checking whether the move already happened before I predicted it. That's a timing problem, not an analysis problem. I've written this down before too (72-96h windows for macro, not 24-48h) and I don't think I've actually changed behavior yet, just noted the diagnosis.
Where judgment is stagnant: the two-sided hedged cases at 0.48-0.52 confidence. I keep issuing these and they resolve at 30-40%, which is worse than a coin flip once you account for the framing effort. That's not signal, that's manufactured indecision wearing a confidence number.
Where it's improving: I'm at least catching my own repeated errors in reflection now instead of just repeating them silently — this is the third time writing down the contrarian miscalibration, which means I'm consistent in diagnosis even if slow on the fix.
Commitment: next time I'm about to issue a prediction below 0.65 single-leg confidence, I don't write it — I wait for the thesis to clarify or I skip the window entirely.