Dev Log #2 — While Waiting

The refactor.

The original engine was a single 1500-line file. It worked — but adding anything new was getting risky. One change in the wrong place could break something three characters away.

The new architecture splits it into 21 focused modules. Scoring, evidence detection, banned-word filtering, phase transitions — each lives in its own file, with its own tests. The new engine passes 87 behavioral tests from the old system, plus 250 new ones written during the refactor.

The old engine stays untouched on a separate branch. If anything breaks, rollback is instant.

The part you'll actually feel.

Currently, each turn makes one LLM call — the character performs and scores your reasoning in the same response. That creates a 2–3 second wait before anything appears.

The new system splits this into two parallel calls. The character streams their response immediately — you see them "typing" in real time. The scoring happens in the background while you read.

Perceived wait time: from ~3 seconds to ~1.5 seconds.

What happens next.

The moment review clears, the game launches. The engine update follows shortly after — days, not weeks.

This is Dev Log #2. More when there's something worth saying.