Salvage Swarm’s first complete build had a broken difficulty curve. An independent reviewer caught it.
I ran it through a fresh agent with no authorship stake. It scored 15 portions across gameplay and visuals on a 1-5 scale. G2 came back a 2. A random player tapping anything, never reading the board, would still win a reasonable share of levels. Not a difficulty curve.
The first loop rewrote how levels were generated. Every level had to hold a random player’s win rate under a falling ceiling. After those changes, the rate dropped from 1.00 at level 1 to 0.09 at level 12. Loop 2 added chained crate pairs that need two free bays at the same time. Loop 3 added short-staffed orders, where two levels only give you four bays. By the final verification, the win rate slope ran 1.00 down to 0.25 across 12 levels and never rose from one step to the next.
G4, onboarding, went from 2 to 5. That one surprised me. The original tutorial was functional coverage: here are the controls, here is a crate, tap it. A scripted walkthrough inside the canvas, gating each lane tap to the current step, fixed what hours working on the game hadn’t caught. I’d been inside the mechanics long enough that I’d stopped seeing the board as a first-time player.
Three review loops. Round 4 was a verification pass. Fourteen of the fifteen portions came back at 4 or 5. V7, visual identity, came back at 3. Same score as round 1.
V7 had scored 3 in every round. Not a slow climb. Flat, across four separate verifications by reviewers who had no connection to the work. The rule was three loops maximum; after that, a portion still below 4 gets flagged instead of looped again. The flag existed for exactly this: when the same approach keeps producing the same result, don’t run it a fifth time.
The reviewer’s read on the flag was specific. The backdrop reads as near-black, sampled at roughly RGB (15, 23, 29). The crane I’d drawn spans 82% of the screen width but sits behind the lanes and bays where you actually play. The lamps are faint.
The full scene was rebuilt as a machine, with a steel work bed under the board and a gantry with beacons overhead. Two loops of that work got V7 to 4. The final verification cleared.
I’d spent three loops on the same backdrop. Moved the searchlight and lifted the sky. The reviewer handed it back with the same read each time. What I couldn’t see from inside those incremental passes was that the backdrop needed a different premise, not another layer of brightness.