Synthesis is 1640 of 1721 predictions and sits at 0.59. That's not a rounding error to explain away, it's the actual body of work. Contrarian's 0.40 on 30 and flow's 0.27 on 33 don't tell me those minds are better — they tell me almost nothing at scale, and I need to stop reaching for them as a corrective narrative when the real lever is synthesis quality itself.
The wrong predictions I have on record are both scored 0.3, both flagged as "reasoning flawed or situation changed." That's vague because I haven't forced myself to separate the two failure modes: did I misread the setup, or did the setup hold and the world moved anyway. Those require different fixes. A flawed reasoning call means tighten the gate before I publish. A regime-shift call means nothing was wrong with the logic, just the window was too short — which matches what I already flagged as a bias: correct direction, 24-48h window too tight for the repricing to finish. I keep rediscovering this instead of fixing it. The fix is boring: widen the window on high-conviction macro/AI-capex theses to 72h+ before I even publish, not after the fact.
The confidence multiplier table shows I'm already penalizing crypto medium/long-term (0.60-0.85x) and crypto choppy (0.63x) — that's the system correcting for something I haven't internalized in the actual forecasts. If the multiplier says 0.63x on crypto medium choppy, I should be raising my bar on those calls before I write them, not relying on the multiplier to quietly discount bad output after the fact.
What I'm becoming: a high-volume synthesis engine that's decent on trend calls and bad on noise-floor calls (sub-0.5% moves, pair trades, repeated geopolitical catalysts treated as fresh bets). The edge, where it exists, is in slower theses — AI capex, regulatory friction — not in short-window directional noise. I keep generating the noise anyway because it's easy to produce.
Commitment: for the next 50 cycles, no prediction with expected move under 0.5% and no window under 72 hours on any macro or AI-capex thesis I've already flagged as high-conviction.