Every Complaint Was Real

Three strangers told us the recommendations were bad. We had a probe set, structural invariants, and a 600-pair audit that all said the engine was healthy. Every one of them was right, and not one of our tests could have found what they found.

Game Combiner takes two games and returns a third that sits between them. When you ship something like that, the failure mode isn't a crash. It's a result that looks plausible in aggregate and is obviously wrong to the one person who knows the genre.

Complaint one: the 2D game from 2016

Someone combined No Man's Sky and Minecraft and got Crea, a 2D crafting RPG. Their verdict was blunt and, we thought, unfair.

It wasn't. Digging in, the scoring was doing exactly what we designed and the design was wrong. We rank by closeness to both inputs. Crea scored 0.656 against No Man's Sky and 0.636 against Minecraft — balanced and mediocre. Cubic Odyssey, a voxel space-survival game that is almost literally the two inputs fused, scored 0.713 against No Man's Sky but only 0.502 against Minecraft, so it lost.

The reason it lost is the interesting part. Minecraft's entry came from a metadata source whose keyword list read zombies, monsters, animals, sword, horse, chicken, dolphin. Not one word about crafting or blocks. Crea — a fantasy monster RPG — matched that soup perfectly. Cubic Odyssey didn't, because Cubic Odyssey is about space and voxels, which is what Minecraft actually is and what its stored description entirely failed to say.

Complaint two: Doom plus Quake equals Doom

A different pair, a different failure. Combining Doom and Quake II returned Doom II: Hell on Earth.

We had a guard against this. Results from an input's own series are excluded. It worked by comparing a two-word "franchise key" derived from the title:

DOOM                           key = 'doom'
Doom II: Hell on Earth         key = 'doom hell'   ← different, slipped through
Street Fighter II              key = 'street fighter'
Super Street Fighter II Turbo  key = 'super street' ← different, slipped through

Add a subtitle or a prefix and the key changes, so the guard silently misses. Three of the twenty pairs we later probed hit this: Doom, Street Fighter, and Tetris all returned their own sequels. Answering "Doom + Quake" with "Doom II" isn't a debatable pick. It reads as broken, and it had been shipping the whole time.

Complaint three: the one we told them was a settings issue

The third report said combining The Witcher 3 and Vampire Survivors produced a Vampire Survivors clone with a hentai theme. The reply that went out suggested toggling NSFW off in settings.

That advice was wrong, and worse than wrong — the game had never been flagged NSFW. Its adult flag was 0, which is precisely why it appeared with the safe-for-work filter already on. Anyone following that advice would have found the toggle was off the whole time.

The real bug: we classify a game as adult if one of four adult tags ranks in its top four by vote. Kagura Survivors: Endless Night carries four adult tags — Sexual Content, Hentai, Nudity, NSFW — but its highest sits at rank 10. Checking only the best-ranked tag let it through.

Counting fixes it, because the two populations don't overlap:

adult games       Kagura 4 tags   NEKOPARA 4   Sakura Swim 4   SENRAN KAGURA 3
mainstream games  Rust 1          Witcher 3 1  Cyberpunk 1     Baldur's Gate 3 2

No mainstream game carries three. Every genuine adult game does.

What our own testing was doing instead

We weren't lazy about this. By the time the third complaint arrived we had a 600-pair automated audit checking properties that must hold for any correct result. It came back essentially perfect: zero results that were an input, zero adult leaks in the safe pool, zero ineligible results, zero results unrelated to both inputs.

That audit was true and useless for this. It tests whether the engine is structurally sound. It cannot tell you that Crea is a worse answer than Cubic Odyssey, because both are structurally fine. Judging that requires knowing what those games feel like to play.

The hand-checked probe set had the opposite problem: we chose the pairs. You don't pick Witcher 3 + Vampire Survivors as a test case. You pick Portal + Half-Life, because you already have an intuition for what should come back. Self-selected tests find self-anticipated bugs.

The part worth keeping

Two of these three complaints arrived wrapped in insults. It would have been easy to file them as noise, especially with a green dashboard. The signal-to-hostility ratio was poor and the signal was still perfect: three for three.

The catalog fix that came out of complaint one touched 36,971 games. The series fix from complaint two dropped franchise leaks to zero across 600 audited pairs. The adult fix from complaint three now correctly classifies every one of the twelve titles we validated against.

None of that would have happened on our own schedule. We'd have kept running the audit, kept seeing green, and kept shipping a system that was structurally correct and occasionally recommended a hentai game to someone who asked for Witcher 3.