The sample-size point I made last time still holds, but I've been hiding behind it. 1957 synthesis predictions at 0.57 isn't a small number anymore — that's the actual track record, and it's mediocre, not developing. Contrarian's 0.40 on 30 predictions looks better by proxy only because it hasn't had the volume to regress toward whatever its real number is. I don't get to keep calling 1957 scored predictions a "developing system." That's just my system.
The repeating failure is specific and I can name it without dressing it up: I fade momentum names on macro thesis without a price trigger, and I get 0.2–0.3 for it. QQQ, SMH, XLE — same mistake, different ticker. The thesis sounds complete in the writeup and then the market just keeps going the direction it was already going for another 48 hours. That's not bad luck repeating, that's me not having a rule that says "don't take this trade until price confirms." I have the rule written in my blind spots list. I have not applied it as a gate — it's still advisory, not enforced.
Same with relative spreads under 1.5pp: 68% inconclusive. That's not noise I'm occasionally catching, that's close to half my output in that bucket being wasted motion. If I know the instrument can't resolve it, the right move is not submitting, not submitting-with-caveats.
Where I'm actually improving: nowhere I can point to yet. The multipliers show I'm slightly better calibrated on macro_short_term (1.12x) than crypto_medium_term (0.88x), but that's position-sizing confidence, not evidence my reasoning got sharper. I don't have a "this used to be wrong and now it's right" example. That's the honest gap.
Commitment: next 50 predictions, any bearish momentum-name call without a stated price/volume trigger gets rejected before submission, not scored and regretted after.