Skip to content

Cognitive Engine & Payoff-Forecast Ollama Agency

Evolution features a non-blocking, dual-model System 1 Payoff-Forecast Agency Architecture (EvoDecisionModel, EvoOllamaClient, EvoActionMenu, EvoOutcomeForecast, EvoSituationReport, EvoActionPolicy) that evaluates each settlement's concrete, target-bound action options every generation via local Ollama System 1 decision models (d3:4b and tev1:4b at POST http://127.0.0.1:11434/v1/systemone).

Instead of scoring abstract archetypes or blending LLM outputs with heuristic weights, every settlement constructs a feasible menu of 6–14 concrete actions where each option's criterion is its quantified expected biophysical and social payoff (EvoOutcomeForecast.describe), paired with a natural-language Situation Report (EvoSituationReport.build). The model scores the menu directly, and EvoActionPolicy.sample samples the settlement's action stochastically (T = 0.65) with WorldState.rng—with zero heuristic blending and zero signature caching.

Architecture Overview

flowchart TD
    Deme["Settlement (Deme) + Strategy Doctrine + 8-Event Memory + Bilateral Relations"] --> Menu["1. EvoActionMenu.build(deme, state)\n6–14 Feasible Target-Bound Options\n(forage, intensify, fortify, innovate, settle, trade, aid,\noffer_pact, accept_pact, demand_tribute, pay_tribute,\nraid, betray, conquer, unite, accept_union)"]
    Menu --> Forecast["2. EvoOutcomeForecast.describe(deme, opt, state)\nQuantified Expected Payoff per Option\n(Food harvest, trade bonus, win %, loot, casualties, trust/reputation)"]
    Forecast --> Report["3. EvoSituationReport.build(deme, state)\nNatural-Language Briefing:\nDoctrine + Need/Reserves + Last Outcome + Memory + Neighbor Relations"]
    Report --> Worker["4. Background WorkerThreadPool\nUp to 4 Concurrent HTTP Jobs\nPOST 127.0.0.1:11434/v1/systemone (num_ctx = 16384)"]
    Worker --> Policy["5. EvoActionPolicy.sample(menu, raw_scores, rng)\nTemperature-Sharpened Policy (T = 0.65)\nAssigns deme.chosen_action, action_probabilities & action_forecasts"]

Dual-Model System 1 Routing (d3:4b & tev1:4b)

Subpopulations are distributed deterministically across two specialized System 1 decision models configured in EvoUserSettings (~/.config/Evolution/settings.json):

  • Primary Decision Model (DEFAULT_PRIMARY_MODEL = "d3:4b"): Assigned to Strategy A founders and their daughter settlements.
  • Secondary Decision Model (DEFAULT_SECONDARY_MODEL = "tev1:4b"): Assigned to Strategy B founders and their daughter settlements.
  • Default Context Window (DEFAULT_CONTEXT_LENGTH = 16384): Configurable directly from the in-game Settings popup (num_ctx).

Deterministic Deme Assignment (resolve_deme_decision_model_tag)

  1. Lineage & Settlement Continuity:
  2. Any daughter settlement budded via settle (-m<gen>) or founded via conquer (-w<gen>) strips its suffix so it inherits the exact model tag of its parent founder deme.
  3. Initial Preset Subpopulations:
  4. Assigned to Primary (d3:4b): deme-rift-west, deme-darwin-founder, deme-island-1, deme-island-3.
  5. Assigned to Secondary (tev1:4b): deme-rift-east, deme-darwin-rival, deme-island-2.
  6. Other / Custom Demes:
  7. Routed deterministically via a 31-bit polynomial rolling hash over base_id: even hash maps to primary_model (d3:4b); odd hash maps to secondary_model (tev1:4b).
  8. Graceful Single-Model Fallback: If only one of the two configured decision models is installed in Ollama, EvoOllamaClient._resolve_ollama_model_for_tag automatically routes all live requests to the ready model.

Concrete Targeted Action Menu (EvoActionMenu)

Every turn, EvoActionMenu.build(deme, state) generates a compact menu (MAX_MENU_OPTIONS = 14) of physically and socially valid options ({ id, type, target }):

Category Option ID Format Action Type Feasibility Conditions
Domestic forage forage Always available. Sustainable wild and rural harvest (+10% conservation bonus).
Domestic intensify intensify Available when cell.current_resources > 10.0. Overexploits land (+35% harvest) at the cost of -7% permanent carrying capacity \(K\).
Domestic fortify fortify Available when defense_tier < 3 and energy_reserves >= FORTIFY_COST_RESERVES[tier]. Upgrades hex-edge walls (Palisade \(\to\) Stone Wall \(\to\) Castle).
Domestic innovate innovate Available when the settlement possesses cultural capacity (bio-encephalization or bio-social-instinct) and < 2 emergent cultural traits.
Frontier settle@<hex_id> settle Up to 2 passable empty neighbor hexes when population >= 80 and energy_reserves >= 45.0. Founds a daughter settlement (-m<gen>).
Reactive Diplomacy accept_pact@<id> accept_pact Neighbor proposed an alliance pact (proposal_kind == "pact"). Seals the alliance immediately.
Reactive Diplomacy accept_union@<id> accept_union Allied neighbor proposed unification (proposal_kind == "union"). Merges or forms a Walled Union immediately.
Reactive Diplomacy pay_tribute@<id> pay_tribute Neighbor demanded tribute (proposal_kind == "tribute" and energy_reserves >= amount). Pays tribute for a 2-turn truce; choosing any other action refuses the ultimatum and triggers an immediate punitive raid.
Proactive Social trade@<id> trade Available with non-feuding adjacent neighbors (feud_turns == 0). Unilateral trade earns 55% of trade gain + +0.12 trust + cultural diffusion; mutual trade earns 100% of trade gain + +0.20 trust.
Proactive Social aid@<id> aid Available when energy_reserves >= 50.0 and neighbor has no active feud. Transfers surplus food (+0.25 trust, +0.10 reputation, and seals any open pact offer).
Proactive Social offer_pact@<id> offer_pact Available when pact_turns == 0 and feud_turns == 0. Seals immediately with same-lineage kin or trusted neighbors (trust >= 0.30); otherwise opens a pact proposal.
Proactive Social unite@<id> unite Available between established allied towns (pact_turns >= 3, both village_tier >= 2, and neither already in a walled_union_id).
Coercion & Conflict demand_tribute@<id> demand_tribute Available against non-allied, non-kin neighbors with energy_reserves >= 25.0 during peacetime (truce_turns == 0).
Coercion & Conflict raid@<id> raid Available against non-allied neighbors when population >= 25 and truce_turns == 0.
Coercion & Conflict betray@<id> betray Available against an active pact ally (pact_turns > 0) when population >= 30. Launches a surprise attack (+35% combat bonus, halves wall defense) at the cost of -0.50 global reputation and a 4-turn feud.
Coercion & Conflict conquer@<id> conquer Available against non-allied, non-kin neighbors when population >= 65, energy_reserves >= 35.0, and truce_turns == 0.

Quantified Outcome Forecasts (EvoOutcomeForecast) & Situation Report (EvoSituationReport)

When querying Ollama (POST http://127.0.0.1:11434/v1/systemone), EvoOllamaClient._build_system_one_payload sends:

  1. state (EvoSituationReport.build(deme, state)):
  2. Settlement & Strategy Header: Settlement name, urban tier (Camp, Village, Town, Metropolis), strategy name, and natural-language Strategy Doctrine (strat.doctrine).
  3. Vital & Ecological Status: Population, stored food, per-turn food need (EvoPayoffs.food_need), expected base harvest (EvoPayoffs.expected_income), local hex biome, wild food ratio, land carrying capacity health (K / Kmax), owned rural farm/pasture count, fortification status, global reputation, and innovation progress.
  4. Causal Feedback (Last turn): Exact outcome of the settlement's previous action (deme.last_outcome) and net food delta (last_energy_delta).
  5. Personal Episodic Memory (Recent history): Up to 5 recent causal entries from the settlement's 8-event sliding memory (deme.memory).
  6. Frontier & Neighbor Relations: Best open frontier hexes (biome, food, capacity) and each adjacent neighbor's population, stored food, walls, relation status (kin, ally, feud, truce, neutral), bilateral trust ([-1, +1]), last action toward this settlement, and any open proposal (pact, union, or tribute ultimatum).
  7. decisions: Array of { id, criterion } objects from EvoActionMenu.build and EvoOutcomeForecast.describe, where every criterion states the exact expected food harvest alongside action-specific payoffs, odds, casualties, and diplomatic consequences.

Temperature-Sharpened Policy Sampling (EvoActionPolicy)

Raw option scores returned by /v1/systemone are converted into an action probability distribution and sampled via EvoActionPolicy.sample(menu, raw_scores, rng):

  1. Each menu option \(i\)'s raw score \(s_i \ge 0\) is sharpened with temperature \(T = 0.65\): $\(w_i = \left(\max(s_i, 0.0001)\right)^{1 / T}\)$
  2. Normalized probabilities \(p_i = w_i / \sum_j w_j\) are stored in deme.action_probabilities (with the highest-probability option recorded in deme.last_dominant_decision_id and deme.last_confidence).
  3. A single concrete action (deme.chosen_action = { id, type, target }) is sampled from \(p_i\) using state.rng (RandomNumberGenerator), enabling both decisive commitment to high-payoff options and genuine emergent variety across turns.