nishi code wiki / research / npc minds

NPC intelligence, behaviour and social simulation

SOTA census · compiled 2026-08-01 · 7 axes · ~45 sourced claims · 8 declared gaps

lineage: forked from research_rtrender parent domain: embodiment target: first-person farming-life village sim on the sovereign voxel engine

Claims are labelled SHIPPING / RESEARCH / ANNOUNCED / ABANDONED. Gaps are declared UNVERIFIED rather than guessed.

1. Classical shipping SOTA — the 25-year canon

The pattern across all five healthy systems: integer/enum state, utility scoring, and a legible ledger — no neural nets anywhere in the shipped canon. The Sims' inversion (objects advertise, agents score) is the single highest-leverage trick: content scales by ADDING OBJECTS, not by editing agent code. Shadows of Doubt is the existence proof that hundreds of fully-scheduled citizens fit in a real-time budget if simulation fidelity follows player proximity.

2. Social simulation — the gossip lineage nobody shipped past

The academic lineage: Comme il Faut (McCoy, Treanor, Mateas, Wardrip-Fruin — ojs.aaai.org/index.php/AIIDE/article/view/12454, AIIDE 2011), a rule-based playable social model, shipped as Prom Week (dl.acm.org/doi/10.1145/2159365.2159425, FDG 2011; game released 2012). Then Talk of the Town (Ryan, Summerville, Mateas, Wardrip-Fruin, “Toward Characters Who Observe, Tell, Misremember, and Lie” — EXAG @ AIIDE 2015; jamesryan.world/talktown): characters form knowledge from observation, propagate it through discrete conversations, misremember, forget, and lie — information flow is agent-driven, not abstractly diffused. Open-source successor: Neighborly (Johnson-Bey, Nelson, Mateas — IEEE CoG 2022; github.com/ShiJbey/neighborly), a ToT-style emergent-narrative sandbox. All four: RESEARCH.

3. LLM NPCs — the honest ledger, 2023–2026

Read the ledger honestly: the survivors are small and on-device (inZOI's 0.5 B), the cloud voice/character vendors died or pivoted to local, the highest-profile cloud NPC was jailbroken in hours, and the studio with the biggest budget shipped nothing in two years of demos. Cloud LLM NPCs failed on cost, on safety, and on latency — simultaneously.

4. Sizing the local model — and the hybrid canon

The shipped precedent is exactly one data point: 0.5 B parameters on-device (inZOI, 2025-03). The emerging consensus architecture — visible in inZOI, in Inworld's pivot, and in the research line of §5 — is the hybrid canon: classical utility AI at frame rate, the LLM consulted at low frequency (dialogue moments, day-boundary planning, reflection), never in the per-tick loop.

Against our substrate: the sovereign integer LLM path runs a 0.5 B model at ~1.06 tok/s on CPU today (in-house, 2026-08-01) — a 20-token villager line takes ~19 s, which is ambient/asynchronous territory only (generate during sleep ticks, deliver when done). The measured GPU path (362 GB/s DRAM-resident, Q4_K kernel bit-exact vs CPU reference) projects ~42.1 tok/s on a 14 B Q4_K — sub-second short lines and day-tick planning. Both in-house measurements, 2026-08-01.

5. Memory and planning — the cost wall and the 10–100x door

Generative Agents (Park et al., arxiv.org/abs/2304.03442, 2023-04-07) is the reference architecture: a memory stream, retrieval scored by recency × importance × relevance, periodic reflection into higher-level beliefs, and recursive day planning. It produced believable emergent behaviour (party invitations propagated, agents showed up) — and it cost thousands of US dollars in tokens for 25 agents over two simulated days (paper's own limitations section; commonly cited as ~$2,000 — see UNVERIFIED #3; forbes.com Joon Sung Park interview, 2024-02-20). RESEARCH.

The reductions came fast: Lyfe Agents (arxiv.org/abs/2310.02172, 2023-10-03) — option-action decision hierarchy, asynchronous self-monitoring, Summarize-and-Forget memory — at 10–100x lower cost than prior generative agents in real-time 3D social scenarios. Affordable Generative Agents (arxiv.org/abs/2402.02053, 2024-02-03) — lifestyle-policy reuse and social-relationship compression, same believability benchmarks at a fraction of the calls. Both RESEARCH.

The transferable lesson: the believability came from the memory architecture, not the model — and cost collapses when the LLM is consulted only at decision boundaries while cached policies replay routine days. For a farming sim, most villager days ARE routine: that is the AGA case in its purest form. The classical ledger of §1 (thoughts, facets, schedules) doubles as the memory stream the retrieval layer reads — build it once, serve both masters.

6. Evaluation — there is no yardstick

No industry-standard NPC believability benchmark exists (as of 2026-08-01; declared, not proven — see UNVERIFIED #5). The closest things: academic believability rubrics (“Towards an Understanding of Character Believability”, dl.acm.org/doi/10.1145/3582437.3582466, FDG 2023-04) and Generative Agents' human-judged TrueSkill comparisons (2023-04). Nobody ships against a number. Consequence for us: define our own gates — a replay-determinism gate (same seed + save ⇒ identical village, LLM lines included), a schedule-coherence gate (every villager reachable at their scheduled place), and a gossip-provenance gate (every belief traceable to an observation or a retelling chain). All three are mechanical, bit-exact checkable, and match the house gate discipline — a believability bar that CAN fail.

7. What this means for the Nishi stack — the ranked ladder

Constraint filter: proprietary language → WASM; integer/fixed-point deterministic core; no third-party libraries; substrate already has 12 NPCs, a 20-bit appearance genome, inherited bold/shy personality bits, WANDER/SOCIALIZE/REST/FORAGE/FLEE states, an entity store, save/load, and bit-exact sim. Rungs 1–5 are classical and cheap; rungs 6–7 are LLM and expensive. Climb in order — each rung's data structures feed the next.

Patent note: the Nemesis-style procedurally-promoted vendetta hierarchy is claimed by US10926179B2 until 2036-08-11. The gossip/provenance design above is mechanically distinct (information propagation, not rank promotion on player-death events), but any future faction-promotion feature should be checked against the claims first.

Declared UNVERIFIED — do not treat as measured