diff --git a/src/multi_swarm/ga/initial.py b/src/multi_swarm/ga/initial.py
new file mode 100644
index 0000000..3d550d9
--- /dev/null
+++ b/src/multi_swarm/ga/initial.py
@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+import random
+
+from ..genome.hypothesis import HypothesisAgentGenome, ModelTier
+from ..genome.mutation import COGNITIVE_STYLES
+
+STYLE_PROMPTS: dict[str, str] = {
+    "physicist": (
+        "Cerca leggi conservative, simmetrie, regimi di scala. "
+        "Pensa in termini di flussi e potenziali."
+    ),
+    "biologist": (
+        "Cerca pattern adattivi, nicchie ecologiche, "
+        "predator-prey dynamics tra partecipanti del mercato."
+    ),
+    "historian": (
+        "Cerca pattern ricorrenti su scale temporali multiple, "
+        "analogie con regimi storici, mean reversion strutturali."
+    ),
+    "meteorologist": (
+        "Cerca regimi di volatilità che si autoalimentano, "
+        "transizioni di stato come fronti, persistenza locale."
+    ),
+    "ecologist": (
+        "Cerca interazioni multi-asset, correlazioni cluster, "
+        "segnali di stress sistemico nelle dinamiche di flusso."
+    ),
+    "engineer": (
+        "Cerca segnali con rapporto S/N favorevole, filtri causali, "
+        "robustezza a perturbazioni di calibrazione."
+    ),
+}
+
+
+def build_initial_population(
+    k: int,
+    model_tier: ModelTier,
+    rng: random.Random,
+    feature_pool: tuple[str, ...] = ("close", "high", "low", "volume"),
+) -> list[HypothesisAgentGenome]:
+    """Costruisce una popolazione iniziale K varia per stile cognitivo + parametri."""
+    population: list[HypothesisAgentGenome] = []
+    for i in range(k):
+        style = COGNITIVE_STYLES[i % len(COGNITIVE_STYLES)]
+        n_features = rng.randint(1, len(feature_pool))
+        feats = sorted(rng.sample(feature_pool, k=n_features))
+        g = HypothesisAgentGenome(
+            system_prompt=STYLE_PROMPTS[style],
+            feature_access=feats,
+            temperature=round(rng.uniform(0.7, 1.2), 2),
+            top_p=0.95,
+            model_tier=model_tier,
+            lookback_window=rng.choice([100, 150, 200, 300]),
+            cognitive_style=style,
+        )
+        # Seed per garantire id univoco se duplicato (raro ma possibile)
+        while any(g.id == p.id for p in population):
+            g = HypothesisAgentGenome(
+                system_prompt=g.system_prompt + f" [seed-{i}-{rng.randint(0, 1_000_000)}]",
+                feature_access=g.feature_access,
+                temperature=g.temperature,
+                top_p=g.top_p,
+                model_tier=g.model_tier,
+                lookback_window=g.lookback_window,
+                cognitive_style=g.cognitive_style,
+            )
+        population.append(g)
+    return population
diff --git a/tests/unit/test_ga_initial.py b/tests/unit/test_ga_initial.py
new file mode 100644
index 0000000..00e11d8
--- /dev/null
+++ b/tests/unit/test_ga_initial.py
@@ -0,0 +1,27 @@
+import random
+
+from multi_swarm.ga.initial import build_initial_population
+from multi_swarm.genome.hypothesis import ModelTier
+
+
+def test_initial_population_size():
+    pop = build_initial_population(k=20, model_tier=ModelTier.C, rng=random.Random(0))
+    assert len(pop) == 20
+
+
+def test_initial_population_unique_ids():
+    pop = build_initial_population(k=20, model_tier=ModelTier.C, rng=random.Random(0))
+    ids = {g.id for g in pop}
+    assert len(ids) == 20
+
+
+def test_initial_population_covers_all_styles():
+    pop = build_initial_population(k=12, model_tier=ModelTier.C, rng=random.Random(0))
+    styles = {g.cognitive_style for g in pop}
+    assert len(styles) == 6
+
+
+def test_initial_population_generation_zero():
+    pop = build_initial_population(k=20, model_tier=ModelTier.C, rng=random.Random(0))
+    assert all(g.generation == 0 for g in pop)
+    assert all(g.parent_ids == [] for g in pop)