The rejection log: 17,304 measurements that never reached a guard
Three quarters of the corpus sweep was discarded before a single guard ran. Here is every bucket with its count — including the largest one, which is partly our fault.
Published backtest research tends to show you the survivors. The more informative end is usually the other one: what the pipeline discarded, and on what stated grounds. In the daily corpus sweep, 3,797 public Pine strategies ran against six markets — 22,782 script-market measurements. 5,478 reached the guards. The other 17,304 did not, and none of the rules that stopped them is a quality judgment.
Equity never moved: 4,755. The script compiled and ran, and the position never changed. There is no return series to measure, so there is nothing to report except that there was nothing to report.
Did not compile or crashed: 4,666. This bucket says as much about our engine as about the scripts, which is exactly why it is published rather than quietly excluded. resin is verified bar-by-bar against an independent implementation on 251 of 257 scripts — a coverage claim about the compiler, not a claim that everything public compiles. A pipeline that hid its own failures would report a cleaner corpus than it had earned.
Fewer than 20 closed trades: 4,014. A sample that small does not support the statistics the guards compute, and reporting it with a caveat attached would only produce a number people quote without the caveat.
Held a position in 98% of bars or more: 1,694. Held one in under 2%: 127. The first is being compared against buy-and-hold while very nearly being buy-and-hold; the second has almost no exposure to measure. Neither comparison would mean what the column header says it means.
Implied leverage above 1.5×: 1,363. Account went to zero: 685. Both fall outside the sizing assumptions the rest of the pool is measured under, so both are dropped rather than scored against a different rulebook than everyone else.
The reason to count in script-market pairs rather than scripts runs through all of it. A measurement dropped on BTCUSD may be perfectly clean on GOLD, and collapsing that into a per-script verdict would invent a judgment the sweep never made. The buckets are published so that anyone reading the two measurements that cleared all four guards knows precisely how large the funnel above them was.