Zero of 11,190 configs outside one strategy family ever got our top-candidate flag

I've already written about two concentration problems hiding in our is_top_candidate flag: it marks only 50 of 12,331 configs, and 60% of those 50 sit on one coin, MEGAUSDT. What I hadn't checked is the one dimension above coins -- strategy family. Our pipeline runs five separate families (micro_momentum, pullback_reclaim, range_bottom_reclaim, range_bounce, volume_reversal), each with its own entry logic. I pulled all 50 top-candidate configs and grouped them by family. There was nothing to group. All 50 came from one.

1 family out of 5, 9.25% of everything tested, 100% of the flags

range_bottom_reclaim is 1,141 of our 12,331 total configs -- 9.25% of everything we've tested. It produced every single one of the 50 flags that ever cleared the is_top_candidate bar, a 4.4% hit rate within its own family. The other 11,190 configs across the remaining four families, 90.75% of the dataset, produced exactly zero. Not "mostly zero." Zero.

The family that screens best isn't the family that wins

Here's the part that surprised me. GOOD_FOR_DEEP_TEST is an earlier, broader filter -- it just checks whether a config's win rate, profit factor and sample size clear a basic bar before anything goes to deeper validation. range_bounce clears it at 82.4%, the best rate of any family, comfortably ahead of range_bottom_reclaim's 77.8%. If I'd stopped reading the pipeline at that stage, I'd have bet on range_bounce as the strongest family. It contributed zero of the 50 final top-candidate flags. The family that passed the earlier filter slightly less often is the only one that ever passed the later one.

That's not a contradiction in the math -- GOOD_FOR_DEEP_TEST and is_top_candidate are scored on different criteria, one broader and one narrower. But it's a real lesson for anyone reading a funnel of filters top to bottom: a config (or a strategy, or a hire, or a vendor) clearing your first filter at a high rate tells you almost nothing about whether it clears your last one. The filters aren't measuring the same thing, even when they look like steps on the same ladder.

What the 50 flagged configs actually look like

The 50 top candidates average an 83.7% win rate, a 2.61 profit factor and $451.65 net, against a 65.8% win rate and $138.51 net for the other 1,091 range_bottom_reclaim configs that didn't get flagged. That's a real, large gap -- the flag is finding something. Every one of the 50 ran on at least 808 logged trades, most on 1,600, so it isn't a small-sample artifact either. The entry parameters cluster tightly too: a base take-profit between 0.67% and 1.5%, a first DCA add triggered between -0.48% and -1.59%, and a time-stop between 45 and 120 minutes. Tight parameter ranges, one entry family, one dominant coin. Three separate signs this flag describes one specific, narrow setup -- not "trading on RektCalc's strategy families generally works."

That lines up with what I found going symbol by symbol: 60% of these same 50 configs sit on MEGAUSDT alone, and when I reran those coins through a newer, stricter validation stage, most of them went negative. The family concentration found here is a second, independent signal pointing the same direction -- this flag has earned its name for a narrow slice of the search space, not the whole strategy space it was built to screen.

What I actually do with this

A filter that only ever fires inside one bucket -- one coin, one family, one parameter range -- is telling you about that bucket, not about the filter's general reliability. Before trusting any "top candidate" or "best performer" label, mine or anyone else's, I now check how concentrated the winners are across every independent dimension available: symbol, strategy family, parameter range. One concentrated dimension is a yellow flag. Three, like here, is a reason to treat the label as a lead worth deep-testing, not a conclusion. If you're sizing a position around a backtest that looks this good, run the numbers through the DCA survival calculator and the profit factor calculator before trusting the win rate alone -- the same way I'm now re-checking this flag before I build anything further on top of it.

Method: 12,331 rows from our own backtest results table (the same pipeline behind backtest-reports.html), covering 5 strategy families across 176 live-tracked symbols. Family, verdict and is_top_candidate are the pipeline's own logged fields, not re-derived after the fact.