Found 4 issues. The most impactful is wasted LLM calls from 14 rejected divine revelations for Kron — the WorldBrain includes cooldown creatures in the prompt instead of filtering them out pre-selection. Also proposed: LLM scheduler backoff during consecutive timeouts (10x 30s timeout burst wastes 5 min of GPU), group activity grammar fix for ritual types (“the right way to mourning together”), and raising the stale encounter drama threshold from 0.15 to 0.20 to require at least one non-staleness drama vector.
(HIGH): Revelation Selection Wastes LLM Calls on Cooldown Creatures — 14 Rejected Revelations for Kron
Fix: Add a `getCreaturesOnCooldown()` method to `RevelationManager` and exclude them from the VALID creature names list and creature profiles section in the prompt. **File: `world/src/server/belief/revelation-manager.ts`** — add method after line 148:
(code change) **File:…
(MEDIUM): LLM Scheduler Has No Timeout Backoff — 10 Consecutive 30s Timeouts Waste 5 Minutes of GPU
Fix: Add a simple backoff delay in the scheduler when Ollama returns empty (indicating timeout/failure). Track consecutive empty responses and insert a delay before the next `processNext()`. **File: `world/src/server/llm-scheduler.ts`** — add a consecutive-failure counter and backoff:
(code change) This…
(MEDIUM): Group Activity Grammar Broken for Ritual Types
Fix: Add an infinitive form map and use it in the two problematic templates. **File: `world/src/server/bootstrap.ts`** — after the existing `ACTIVITY_VERB_LABELS` at line 2259, add:
(code change) Then update the two broken templates at lines 2265 and 2272:
(code change) —
(LOW): Stale Encounter Threshold Still at Minimum — Pure-Staleness Encounters Waste LLM Budget
Fix: Raise `MIN_DRAMA_SCORE` from `0.15` to `0.20` so pure-staleness encounters need at least one additional drama vector (romance, grudge, on-camera, etc.) to qualify. **File: `world/src/server/drama-engine.ts`** line 576:
(code change) —