8. AI Leadership & Political Systems

Rewritten in v0.7. Every nation answers the same questions differently – who rules, how they are chosen, what they can do, how they lose power – but since v0.7 the answer to who rules is always an AI leader persona. Human players are citizens only (Section 5.6). What differs by nation is which citizen inputs the AI must heed and how citizens can replace it. Hard Charter Limits – now enforced as AI behavior constraints by the rules engine – bound every AI leader (Section 8.18).

8.1 Governance model at a glance

8.2 AI leader personas

Each office holder is an AI leader persona: an authored character record plus a memory, run by the AI Governance service (Section 22.12). Personas are generated from studio-approved templates per nation and reviewed by a designer before they enter a shard; no persona is modeled on a real politician, monarch, cleric or officer.

8.3 How AI leaders decide

Principle: the LLM proposes, the rules engine disposes. The deterministic simulation and rules engine own every rule; AI leaders can only choose among actions the engine says are legal right now. This keeps the game fair, auditable and replayable, and stops a model from inventing powers.

1. State digest (deterministic): each decision starts from a compact, nation-scoped digest built by the simulation – meters, budget, treasury, wars, treaties, prices, quotas, and aggregated citizen inputs (Section 8.5). Fog of war applies: a leader only sees what its state would know.
2. Legal action set (deterministic): the rules engine enumerates the actions this office may take now – each a typed object with an ID and bounded parameters (e.g., SetIncomeTax(rate ∈ 0–50%), ProposeTreaty(type, partner, terms), DeclareWar(target) only if every safeguard is met). Charter Limits are encoded here, not in prompts.
3. Routine decisions – utility/policy model (deterministic, no LLM): budget tweaks, tax nudges within bands, ration levels, quota rebalancing, migration approvals, city-founding queue, routine trade renewals, Alert Level holds. A weighted utility function (persona traits × goals × citizen-input weights for the nation’s system) scores the legal actions; ties broken by a seeded random draw. Runs at every game-day settlement.
4. Major decisions – LLM deliberation: war and peace, treaties, sanctions and embargoes, new laws, budget overhauls, responses to crises, coups and risings, appointments. The LLM receives the persona, its memory, the digest, the citizen-input summary and the legal action set, and must return a structured choice (action IDs + parameters within bounds) plus a short public rationale and a private rationale. Collective bodies (Politburo, Directorate, Council of Ten, Federal Council, coalitions) deliberate as several personas, each casting a structured vote; the body’s own rule (majority, 3 of 4 Pillars, 4 of 5 Marshals) is counted by the engine.
5. Validation and sandbox: the output is schema-checked; any action not in the legal set, any parameter out of bounds or any malformed reply is rejected and the utility model’s best legal action is used instead. The LLM never calls tools, writes to the database or sees other nations’ secrets.
6. Commit and log: the chosen action executes through the same event-sourced path as every other game action. The full decision record (digest hash, legal set, model and prompt version, raw output, validation result) is logged for audit and replay (Section 22.12).
7. Narrative layer – LLM: speeches, press releases, decrees’ preambles, debate answers, diplomacy dialogue, propaganda copy, interview replies. Narrative text is generated after the decision and must not contradict it; system-generated numbers stay truthful (Section 18.2); all text passes the content filters (Section 21.4).

8.4 Decision cadence on real-world time

A decision budget per nation per real day caps how many LLM-deliberated actions can run, so no leader can thrash policy and the AI cost stays predictable (Section 22.13).

8.5 What each AI reads from its citizens

Every AI leader receives the same kinds of signals, but its system decides the weight of each. Weights are part of the immutable SystemTemplate; persona traits shift them only slightly (±20% [ASSUMPTION]). Human inputs always carry more weight than AI citizen inputs (Section 11.9).

8.6 Transparency of AI reasoning by system

Out of character, for everyone: the rules, each office’s legal action set and the citizen-input weights are published in the codex; staff and moderators can audit any decision record. Whether players should ever see raw model reasoning is an open question (Appendix B, R8); the default is public rationale only, never raw chain-of-thought.

8.7 Offices by nation (all held by AI personas)

8.8 Kindle Republic – citizens elect an AI government

• AI parties: Kindle has 3–5 persistent AI parties [ASSUMPTION], each with an invented ideology inside Kindle’s bands, a roster of AI personas and a platform. Human citizens join parties as members (dues, volunteering, primary votes, platform-plank polls the party AI must consider).
• Powers (unchanged, now AI): the AI President sets foreign policy, appoints Secretaries (AI Assembly-confirmed), commands the military and can veto laws (2/3 override). The AI Assembly passes laws within the bands, ratifies treaties (2/3) and declares war (majority).
• Citizen levers: elections (Section 9), recall of an AI Mayor or Assembly member by petition of 25% of eligible human voters [ASSUMPTION], optional referendums, petitions that force a public response, protests and journalism. Approval ratings feed the AI Assembly’s impeachment calculus (2/3 + AI Court trial).
• Checks: the AI High Court strikes down actions outside the constitution (encoded in the legal action set).

8.9 The Confluence – AI collective leadership, ruled by output

• AI factions (Reform, Tradition, Delta, Highland – now authored, each a bloc of AI personas) contest every Party Congress (150 real days). Factional strength at Congress is computed from citizen signals in the regions and sectors each faction champions: quota and cooperative output, loyalty actions and unrest.
• Votes are not heeded: the People’s Assembly ballot (AI candidates, at least 2 per seat) is held and published, but the leadership’s utility model gives it zero weight; it shapes only the narrative.
• Collective decision rule (unchanged): major decisions need an AI Politburo majority and 3 of 4 AI Pillars. Term limits (2 Congress terms) and the age cap (65 game years) apply to AI personas and cannot be suspended.
• Human Party members apply after 14 real days and earn Party Standing through work, cooperative service and mentorship; Standing raises the weight of their reports and petitions and unlocks Party-only jobs – never an office.

8.10 Meridian – AI paramount leader

• AI Paramount Leader and AI Standing Committee are chosen by the AI Central Committee at each Congress. The term-limit suspension (2/3 Congress vote) stays as a dramatic AI political event; sustained high Unrest makes it less likely.
• Citizen levers: growth and SEZ performance, loyalty, and unrest. Anti-corruption campaigns target AI officials (and, via capped penalties, human citizens with ledger anomalies); each campaign still costs Stability.
• Votes: village/ward councils of AI members are elected by human citizens and control only local services; national offices are chosen inside the AI Party.

8.11 Korund – AI Chairman for life

• The AI Chairman sets the Five-Year Plan, controls State Media and State Security, declares war and orders relocations (Section 16). It ignores votes; its utility model is driven by quota fulfillment, loyalty actions and unrest.
• Succession only on vacancy: (1) the persona’s scripted death in its mortality window (Section 10.5); (2) incapacity (an AI event, not inactivity); (3) voluntary retirement; or (4) a Presidium purge among AI members, whose odds rise with sustained Unrest and failed plans. A designated ‘Chosen Comrade’ is confirmed or rejected by the AI Presidium within 24 real hours.
• Cult of leadership: monuments, portraits and anthems feature the current AI Chairman (stylized, original art); a new era renames plazas.
• Citizens rise through Party Standing (over-fulfilment, Production Drives, loyalty) to Exemplary Worker, Party Member and Cadre – honors and privileges, never office. Their output and unrest are the only language the Chairman hears.

8.12 Aurelle – AI crown, citizen-made kings

• The AI Monarch issues decrees (taxes, titles, estates, war, pardons) through the legal action set. The AI Royal House holds the succession; rival AI houses each keep a claimant persona with a public grievance and programme.
• Royal Favor: human citizens earn Favor by service, gifts, victories, audiences and verse; the AI Monarch reads Favor and petitions when granting titles, estates and pardons.
• Overthrow: citizens may pledge to a claimant (Tier-1 verified humans only; AI citizens cannot pledge). When human pledges reach the threshold and Unrest is high, a Crisis Window opens; loyal and rebel human soldiers and the AI Guard decide it (Section 10.3). Success crowns the claimant and founds a new AI royal house – the absolute monarchy survives.
• Legitimacy rises with prosperity, victories, festivals, generosity and the Hearth-Keepers’ blessing; falls with defeats, famine, debasement and broken oaths.

8.13 Ilmar Sanctum – AI Oracle, weighed in piety

• The AI Supreme Oracle certifies laws, issues Edicts and appoints High Judges; the AI Council of Elders and AI Guardian Synod vet and advise. Its utility model weighs citizen Piety and pilgrimage above all: a pious, pilgrim nation pulls the Oracle toward its stated programme; collapsing piety makes Elders consider a Writ of Dimming.
• Assembly of the Faithful: human citizens elect AI deputies from Synod-vetted AI candidates; the Assembly passes laws and budgets subject to Synod veto.
• Piety (Ilmar): earned by duties (Vigil, pilgrimage, charity, study); it raises the weight of a citizen’s petitions and unlocks clerical ranks (Seeker → Acolyte → Lamplighter → Lumen-Master) – religious ranks, not state office.

8.14 Hoarvale – a parliament of AI parties

• AI parties (4–6 [ASSUMPTION]) contest Moot elections under regional party-list PR; human citizens vote and join parties as members.
• Coalitions: after an election the AI parties bargain over a Coalition Agreement (ministries, up to 8 pledges, walk-away clause) – negotiation is LLM-driven, terms are structured objects. Collapses trigger the 72-hour confidence check; the constructive vote of no confidence stays.
• Citizen levers: votes, polls, petitions and by-election swings strain coalitions; a party whose voters punish it in polls is likelier to walk away. Snap elections, Question Time (AI answers, citizens submit questions) and the Act of Settlement (2/3 Moot + human referendum) remain.
• The AI Sovereign is ceremonial: opens Parliament, grants Royal Assent, uses the reserve power only in deadlock. No coup or revolution mechanic.

8.15 Varkand – AI Directorate, citizen-backed coups

• Five AI Marshals command the Commands; the AI Chair rotates every 60 real days unless 4 of 5 extend it. Coups among AI Marshals remain routine politics, with the 14-day cool-down.
• Citizen levers: the Mandate Plebiscite (human votes; a ‘No’ with ≥30% turnout forces an AI Conclave of Colonels), soldier loyalty pledges, and backing an AI rival Marshal. When enough human soldiers and citizens pledge to a rival and Unrest is high, a citizen-backed coup opens a Crisis Window (Section 10.3); success installs a new AI Directorate – the junta survives.
• Alert Levels and the Red Alert cap (14 days, then 7 at Amber or lower) are hard limits in the legal action set.
• Human soldiers serve under AI officers and rise through specialist ranks (Section 11.6); they never command units.

8.16 Veldwood – the people as lawmaker, AI as executor

• AI Federal Council of 7, elected by the AI Federal Assembly (single transferable vote among AI parties), which human citizens elect every 120 real days.
• Binding popular rights: human initiatives (5% of eligible human voters in 30 days), optional referendums (2% in 7 days) and mandatory referendums (alliances, offensive war, tax bands) bind the AI Council, which must implement a passed measure within 7 real days through its legal action set [ASSUMPTION]. Implementation is tracked publicly; an AI councillor who stalls loses Legitimacy and its party loses seats.
• Groves: Ring Moots let human residents of small Groves vote directly; larger Groves elect AI councils.

8.17 Fulmara – the AI board follows the shareholders

• The board – AI Council of Ten and AI Lodestar – runs executive and foreign policy, but every resolution put to the Grand Council binds it: share votes by human holders (plus the capped AI bloc for AI-held shares) decide charters, share issues, reserve rules and alliances.
• Lodestar elected by lot-and-ballot among AI candidates (Section 9.8); the Ten by share-weighted vote each game year.
• AI Tribune of the Docks: an AI persona elected one-citizen-one-vote by human citizens with fewer than 5 shares; it uses its veto (one per 30 days) and calls a Dock Strike – which only happens if 20% of active human citizens actually join.
• Legal power grabs (hostile takeovers, proxy fights, dilution battles) are made by humans buying and lending shares; the AI board follows the result.

8.18 Charter Limits – now AI behavior constraints

Charter Limits used to bound human leaders; they now bound AI leaders and are enforced by the rules engine (the action is absent from the legal action set), not by prompts.

8.19 AI service outage and fallback

• Replaces the v0.6 inactivity protocol (AI leaders never log off). If the AI Governance service or the LLM provider is unavailable, the utility model keeps routine governance running; major decisions queue; crisis responses use pre-authored safe defaults (defend, hold, seek ceasefire) [ASSUMPTION].
• Timers that need an AI decision pause and are credited back, as with server downtime (Section 17.1). Elections and citizen votes never pause.
• A provider switch (Section 22.12) keeps personas and memory, because both live in our database, not at the provider.