advisors mechanic sourced
Advisors
In-world advisor characters monitor simulation indicators, summarize problems, and recommend actions.
definition
An advisor layer watches a set of indicators (budget balance, unemployment, crime, service coverage, happiness, military strength) and maps threshold crossings or rule matches to authored messages spoken by a named character, usually with a recommended action. It owns no simulation state: only trigger conditions, a priority per message, cooldowns, and an optional persona per department. Each evaluation interval it tests its conditions, ranks firing messages by severity, suppresses those still on cooldown, and presents the top one as a pop-up, voiced line, or entry in an advisor panel the player opens on demand. Inputs are aggregate state variables; outputs are text and, in some games, a one-click link to the relevant tool. Canonical forms: the department advisors of SimCity 2000 and SimCity 4, the Super NES SimCity advisor Dr. Wright who appears to report city problems, the advisor screens of the Civilization series, and Tropico's Penultimo.
player experience
The player gets a running diagnosis without reading every graph. Early on advisors teach which variables matter and what usually fixes them; later the player weighs advice against their own model, and conflicting advisors make tradeoffs explicit. Bad or repetitive advice is felt as nagging.
design levers
- which indicators each advisor watches and the threshold or trend that triggers a message
- diagnosis only versus a concrete recommendation or a one-click fix
- one generalist advisor versus a council of department advisors with conflicting priorities
- delivery mode: interrupting pop-up, voiced line, or passive panel opened on demand
- cooldown and repetition rules per message
- accuracy: whether advisors read true state, a lagged view, or a view biased toward their department
failure modes
- A message keeps firing while the player cannot act on it, so players mute or ignore the advisor entirely.
- Advice derived from a single indicator recommends the wrong fix when the cause is an interaction, such as more police when crime is driven by unemployment.
- Advisors that are always right reduce the player to following instructions; advisors that are often wrong teach players to ignore all advice.
- Lines written for early-game states become irrelevant or wrong at late-game scale.
dynamics
- Advisors act as an embedded tutorial whose value fades as the player's own causal model improves.
- Department advisors arguing for opposite policies externalize tradeoffs such as tax revenue versus satisfaction.
- When advice is trusted, player builds converge on what the advisor recommends, narrowing strategic variety.
open questions
- Can a language-model advisor grounded in live state queries answer arbitrary why-questions accurately enough to be trusted, and how should it express uncertainty?
- Does advice deliberately biased by an advisor's department or personality produce better decisions than neutral advice?
ratings
12 from named runs; stored scores are their lower medianoriginal score
Named advisors explicitly diagnose indicators and recommend actions, making their role mostly intuitive. Biased views and conflicting departments produce surprises below 5, without players normally needing deliberate tools to understand the message layer (3).
original score
Threshold checks select authored messages without changing simulation state. Player responses can affect other systems, but the advice layer itself supplies no unscripted state variation warranting 2.
original score
Distinct personas, state coverage, recommendations and optional voice require substantial writing and presentation pipelines. This exceeds moderate bespoke variants (3), but systematic triggers and reuse mean content is not necessarily the primary ceiling on depth (5).
original score
Periodic aggregate threshold tests, cooldown checks and ranking a small message set are negligible. The monitored simulations own their own costs, so modest linear runtime overhead (2) is not required at typical scale.
original score
Severity, thresholds, timing, bias and recommendation accuracy interact and need playtesting. This exceeds a few stable parameters (2), but advice selection itself does not routinely create sensitive runaway strategies deserving 4.
original score
Grounded models can diagnose interacting causes and tailor explanations or council debate, making substantially new advisory forms feasible beyond cheaper writing alone (3). Monitoring and recommending remain the defining role, excluding 5.
appears in
the paper, 3 sections · 2 chapters · 1 episodebacklinks
computed from other records
LLM explanation of simulation state
On-device small language models for charactersclaims
1 with sourcesThe Super NES version of SimCity added Dr. Wright, an advisor character who appears to report problems in the city.