21. Moderation, Safety & Sensitivity

21.1 Key risks

Risk	Why it is specific to this game	Mitigation
Real-world political flame wars	Players map nations to real countries and parties.	Clear rule: in-world politics yes; real-world parties, politicians, current events and hate symbols no. Filters + human moderators; OOC topics routed to a moderated off-topic forum.
Hate speech, extremist content	Systems inspired by authoritarian models can attract players who glorify real regimes or atrocities.	Banned symbols list (real-world extremist and regime insignia), no real flags, slogans or leaders; zero tolerance for praise of real atrocities; proactive review of posters and names.
Power-based harassment (revised v0.7)	No human holds power over another any more, but AI penalties (Correction Duty, raids) can still feel personal, and players can still harass each other over politics.	Charter Limits as AI behavior constraints (Section 8.18): rule-triggered, capped, appealable punishments; AI never names or denounces individual players; player-to-player harassment rules unchanged.
AI output turns hateful or extremist (v0.7)	AI leaders write speeches and state media in authoritarian styles.	AI content policy, layered filters, stylized invented rhetoric, human review of flagged output, kill switch (Section 21.4).
AI read as real-world commentary or bias (v0.7)	Players may read an AI leader’s views as the studio’s real-world politics.	Invented ideologies only; no real-world issues, parties or figures in prompts or outputs; bias evals across nations and personas (Section 22.15).
Prompt injection and manipulation (v0.7)	Players try to steer AI leaders or Notables through crafted petitions or chat.	Citizen text enters only as quoted, clustered data; legal action set; Notables cannot affect metrics; attempts logged and moderated (Section 11.10).
Confusing AI with humans (v0.7)	Players may not realize who is AI.	AI badges everywhere, AI-generated labels, disclosure in onboarding (Sections 11.10, 19.1).
In-character vs out-of-character confusion	‘Informing’ and ‘surveillance’ are mechanics.	Mechanics act on game actions and in-character channels only. Moderation reports are a separate, clearly labeled OOC tool; in-game ‘denunciation’ is a different button.
Minors	Open chat, politics and coups.	Minimum age 16 (decided v0.4; no under-16 accounts or 13+ variant), enforced by a neutral age screen at onboarding (Section 22.3); chat filters on by default; DMs from strangers limited; report/block everywhere.
Sensitive depictions	Theocracy, communism, military rule and oligarchy inspired by real systems; real believers, citizens of such states and survivors of military regimes may play.	Invented faiths with no borrowed doctrine; no real religious symbols; no real army insignia, salutes or historical coups and massacres referenced in Varkand; satire aimed at power, never at peoples; sensitivity readers from relevant backgrounds review nation and faith bibles.
Invented faiths read as real religions (v0.3)	Players may map the six faiths – or state atheism and religious suppression – onto real religions and real persecution.	Archetype-based design (Section 14); respectful codex; no faith is a villain; real-religion comparisons in chat are handled like real-world politics (off-topic forum); persecution mechanics are capped, never graphic, and never target a player’s real faith.
Religious harassment (v0.3)	Faith becomes a vector for in-group/out-group abuse.	Mute missionaries; invitation limits; slurs against real or invented faiths moderated alike; only the player can change their faith.
Doxxing / real-life threats	High-stakes politics can spill over.	Strict rules; quick escalation; no real names visible; privacy defaults.

21.2 Presentation principles

How we present real-world systems

Each system is a stylized simulation of a structure (how power is selected, how goods are allocated), not a portrait of a country, people, religion or culture.
Every nation gets strengths that its citizens can be proud of and weaknesses that create drama. No nation is designed to be the ‘villain’ or the ‘correct’ choice.
Aesthetics derive from government, culture, climate and geography, currency and invented faiths – not from real ethnic or national dress, scripts, religious symbols or landmarks.
The in-game codex explains each system neutrally, notes that real countries vary widely and change over time, and avoids claims about any real state.
Sensitivity review before reveal: political scientists or civics educators, plus readers with lived experience of several of the inspiring systems [to recruit].

21.3 Moderation operations

Tools: report button on every message, name, poster, profile and every AI-generated speech, article or Notable reply; mod console with context (chat logs, actions, AI decision records); shadow-mute; temporary chat bans; vote and pledge voiding; account suspension; per-persona and global AI kill switches.
Staffing: at MVP, the founder plus 2–4 trusted volunteer moderators with paid part-time lead [ASSUMPTION]; scale with population. Consider a moderation vendor at Early Access [VERIFY cost].
Automated first pass: text filters, toxicity classifiers, image-free UGC (templates only) at MVP.
Transparency: public rules, quarterly moderation summary (including AI-content incidents), appeal process.

21.4 AI content safety (new in v0.7)

Layer	Design
Content policy	AI leaders, candidates and Notables must not generate hateful, dehumanizing or extremist content; praise of real regimes, atrocities or extremist movements; real-world politicians, parties, flags, slogans, events, religions or current affairs; threats or harassment of any player; sexual content; or instructions for real-world harm. In-world rhetoric is stylized and invented: production slogans, royal panegyrics, canon homilies, campaign promises – aimed at invented policies and rival governments, never at peoples, ethnicities, faiths or real groups.
Prompt layer	Each persona’s system prompt carries the content policy, the nation’s style guide and a banned-topic list; prompts are versioned and reviewed like code.
Output filters	Every generated text passes (1) a blocklist of real names, symbols and slogans, (2) a moderation classifier (for example a provider moderation endpoint – OpenAI lists omni-moderation-latest as free, checked Oct 13, 2026 – or an equivalent) and (3) a stricter threshold for public broadcasts. Failing text is regenerated once, then replaced by a templated line; the decision itself is unaffected.
Human review	Flagged outputs and a random sample of public AI texts (e.g., 1% [ASSUMPTION]) go to a review queue; sensitivity readers review persona templates and style guides before each nation ships.
Logging	Every prompt version, model ID, output, filter result and publication is logged with the decision record (retention per privacy policy) for audit and replay.
Kill switches	Staff can mute a persona’s narrative output, freeze a nation’s major decisions (the utility model continues) or switch all narrative to templates – instantly and logged.
Players unaffected	Player-to-player harassment rules, reporting and moderation are unchanged; AI-generated content can be reported like any other.