I logged why an automated filter rejected 895 of 1,728 startup ideas

I build a tool that screens startup ideas for solo founders, and it keeps a record of every candidate it refuses and which check it failed. I had never looked at the whole pile at once. Last week I did, and the distribution surprised me.

33 runs, June to September. 1,728 candidates in, 895 dropped — 51.8%.

The single most common reason, by a distance: 496 ideas were dropped because I could not find a single documented complaint behind them. Not "the market is crowded", not "the economics don't work" — nobody had written the problem down anywhere I could find it. That is 55% of all refusals.

After that it gets more specific:

270 — too much hands-on support for one person
193 — leans too hard on one platform
148 — just an AI wrapper with no product underneath
129 — can't charge enough
109 — hard for a solo founder to sell or run
And the part I keep thinking about: 69% of the dropped ideas failed more than one check. A weak idea is rarely weak in one place. The worst one failed four.

Three things this is not, before anyone asks:

It is not an accuracy claim. I have no idea whether the 895 would have sold. Nobody does. All I can say is that the gate was applied consistently and wrote down its reasons.
It is not a market sample. These runs are overwhelmingly my own accounts — I have had one paying customer. This is a record of a machine's behaviour, not of demand.
The checks are my opinions, encoded. If you think "no documented problem" is a bad reason to drop an idea, that is the conversation I actually want.
The data is public and reusable (CC BY 4.0) if anyone wants to argue with it or reuse it: https://whittleos.com/benchmark — the JSON is at https://whittleos.com/gate-census.json

The question I cannot answer from my own data: for those of you who shipped something that worked — was there a written complaint behind it before you started, or did you find the problem by living it?