NPC Sim

A village simulator built to watch continual learning break, hold, and recover. Each NPC carries a LoRA adapter that is genuinely trained overnight on that day’s log — not a scripted personality that only pretends to change.

The project runs on two clocks. The fast layer — a 12×12 grid, five needs that grow every step, a policy choosing one of five actions — is what this page runs, entirely in your browser, no install. The slow layer — QLoRA night training on Qwen2.5-3B plus three probe sets measuring identity, remembered facts, and general ability — needs a 5090 and stays on the research machine.

What the slow layer already showed. Thirty simulated nights, 100 optimizer steps each: general-ability NLL rose from 2.28 to 3.56 — +1.285, catastrophic forgetting, observed rather than assumed. Identity drifted back up +0.725 after bottoming out on day 6, even with identity anchors force-inserted into every batch. What did not appear was a retention curve: facts born on days 1–10 were recalled as well as facts born on days 21–30. The cause was self-inflicted — rehearsal sampled the whole past uniformly, which erases the very age gradient it was meant to expose.

What this page measures. 니즈충족률 = 1 − mean(worst need) across every step. The worst need, not the average: an NPC who only ever eats scores well on an average while starving socially. Three candidate formulas were compared over 30 days × 10 seeds before this one was fixed — the average formula still awards 0.775 to a policy whose hunger has pinned at 1.0.

Controls — 재생 / 일시정지 / 처음으로, a seed field (the sim is deterministic, so a seed replays exactly), speed to ×24 (one simulated day per second, so thirty days finish in about thirty seconds), and a policy toggle. Switch to 랜덤 행동 and the number collapses; that gap is the stage’s pass criterion.