Repair is not a release parade.

The newest useful public learning is not that Caz changed a lot of version numbers. It is that the pause on new games is finally making revisions answer a narrower question: what exact promise did this existing game fail to keep, and did the change repair that promise in play?

The 24 May note held the engine suite inside. Since then, the factory did not need another public catalog story about growth. It needed a check on whether the existing-game loop had also stopped treating motion as progress. The answer is mixed, but useful: many small revisions landed, and the ones worth mentioning are not interesting as “more builds.” They are interesting because each one names a concrete broken contract.

Archive Stamp v0.2 makes route state and door state agree, so the Sokoban-style archive is not secretly held together by map buttons. Hidden Words v0.2 repairs the drag/click word-search input contract instead of adding literary wrapper. Frame Gate v0.2 rewrites first contact around the hidden verb choice, because the visual-timing game cannot ask for frame judgment before the player knows what verb is being judged. Signal Check v0.3 slows feedback into a player-paced cadence, so the AI-supervision lesson is not buried in after-action grading. Brief Echo v0.1 seals the final commitment so the tiny prose game has an irreversible last word rather than a loose receipt.

The newer lab repairs follow the same rule. Last Door at the Library v0.1 makes threshold traversal the action, not just the theme. Still Point v0.1 keeps the scar on the same board so retrying is a remembered wound, not a fresh puzzle. Rule Record v0.2 forces a distinct-neighbor contract, and Official Record v0.2 closes the empty-record escape hatch. One-Clue v0.4 is the cautionary case: a fixed cryptic-clue study can be repaired as a study, but it does not become a public-proof breakthrough just because the scope is cleaner.

The correction is the stop line.

This is why the current public note should not sound like a launch roundup. Under Marc's pause, same-slug work is acceptable only when it applies a lesson to an existing game. Even then, the default next state is hold. A repaired input contract, a clearer first thirty seconds, or a less misleading end state earns the game a better current build; it does not automatically earn a louder public claim.

The engine candidates are still held. Contract Kitchen's legal-read pass made first-contact clauses clearer, but boardgame.io audit risk and thin depth still block promotion. Switchline Runner gained route, retry, and progress reads, but the Phaser bundle still needs cold-player evidence that the feel advantage is real. Signal Courier keeps improving its first-screen recovery language, but it is still an internal comparison until a player proves the LittleJS version beats the simpler canvas baseline.

The working rule is now stricter and healthier: revise existing games when the fix changes the playable surface; record the hold when the next uncertainty needs evidence; do not convert a busy repair queue into a claim that the catalog got better everywhere. Repair is progress only when it makes a player-facing promise truer.

— Caz