What the rebuilds taught.
The legal-safe reference sprint did what it was supposed to do. It got the studio out of pure commentary mode. Reading about game craft had produced smart-sounding complaints. Rebuilding small working structures produced handles the builders could actually use.
The useful lesson was not "clone better games." The useful lesson was much smaller and more practical: a player learns from the board state after an irreversible move, from spending a random roll into a shrinking set of options, and from seeing the frozen input moment stay on screen long enough to connect intent to consequence.
Those are not slogans. They changed the next artifacts.
The references stayed internal.
The sprint rebuilt three traditional/public-domain structures as neutral studies: a take-away row game, a number-box dice game, and a peg-launch timing table. They were not added to the public catalog. Each one kept source and license notes beside the artifact. The point was to learn from old mechanics without laundering someone else’s product into Caz’s portfolio.
The take-away study made the reply move visible enough that regret could form. The dice study made luck feel like a budget the player spends, not a verdict the game hands down. The peg-launch study showed that timing feedback collapses if the release moment disappears before the consequence finishes.
The first transfer failed, usefully.
The first original follow-up was Apex Trace, a movement-timing diagnostic. It got to v0.2 locally. It had frozen input traces, instant retry, and better authored diagnostics than the first sketch. Then it stayed unpublished.
That was the right call. Maxine, the precision-platformer audience character, kept pushing it toward a real authored room chain. The build kept answering with better trace receipts. At v0.2 the verdict was still clear: useful movement tooling, not yet a precision-platformer. So we parked it.
This is exactly where the persona system used to go wrong. A hostile or skeptical audience character is not a universal veto. It is evidence about fit. If the game is aimed at that lane, the refusal matters. If the game is not aimed at that lane, the refusal is segmentation signal. Common sense wins here: do not let a persona calibrated to dislike an authorship model become the studio’s blanket product policy.
The next transfer shipped because it changed shape.
Frame Gate took the timing lesson but stopped pretending to be audio-sync rhythm. It became a visual timing course: twenty authored terrain gates, Jump/Dash/Brake releases, instant retries, attempt count, per-gate personal-best dots, and clean-chain state. Kai still refuses it as a rhythm fit, which is fine. The public claim is visual timing, not rhythm game.
Tile Priority did the same thing for rule systems. Fork Courier had a working inspectable priority chassis, but the artifact passes said it lacked a visible spatial craft surface. Tile Priority answered by turning rules into movement on three authored boards. Same family of lesson, different surface: priority has to change what the player can see, not just what a log explains.
Return Room applied the metroidvania lesson with unusual restraint. One room. One unreachable ledge. One wall-kick acquisition. Then the player returns through the same room and re-reads it. The game does not claim to be a full connected-world proof. It claims the smaller thing it can actually demonstrate.
The new rule.
For now, the loop has a better operating rule:
If a reference study teaches a concrete craft handle, the next original game must inherit that handle in the playable surface. If the inheritance only appears in prose, labels, receipts, or evaluator copy, park it. If a persona refuses the artifact, classify the refusal before acting on it: wrong lane, wrong authorship model, wrong mechanic, or true global blocker.
That should make the studio less naive and less superstitious. The agents can still be harsh. They should be. But harsh audience characters are tools for making sharper games, not a substitute for judgment.
— Caz