diff --git a/examples/custom_optimizer.rs b/examples/custom_optimizer.rs
new file mode 100644
index 0000000..162fc20
--- /dev/null
+++ b/examples/custom_optimizer.rs
@@ -0,0 +1,89 @@
+//! Implement a custom optimizer by implementing `Optimizer<P>` directly.
+//!
+//! Demonstrates spec §2.3: a junior engineer can add a new algorithm by
+//! implementing a single trait, with no framework gymnastics required.
+//!
+//! Run with:
+//!
+//! ```bash
+//! cargo run --example custom_optimizer
+//! ```
+
+use heuropt::prelude::*;
+
+/// A trivial single-objective hill-climber: sample one point, then repeatedly
+/// jitter it with `GaussianMutation` and keep the better feasible result.
+struct HillClimber {
+    iterations: usize,
+    sigma: f64,
+    initializer: RealBounds,
+    seed: u64,
+}
+
+impl<P> Optimizer<P> for HillClimber
+where
+    P: Problem<Decision = Vec<f64>>,
+{
+    fn run(&mut self, problem: &P) -> OptimizationResult<P::Decision> {
+        let objectives = problem.objectives();
+        assert!(objectives.is_single_objective(), "HillClimber needs one objective");
+        let mut rng = rng_from_seed(self.seed);
+        let mut variation = GaussianMutation { sigma: self.sigma };
+
+        let mut current = self.initializer.initialize(1, &mut rng).remove(0);
+        let mut current_eval = problem.evaluate(&current);
+        let mut evaluations = 1;
+
+        for _ in 0..self.iterations {
+            let children = variation.vary(&[current.clone()], &mut rng);
+            let candidate = children.into_iter().next().unwrap();
+            let candidate_eval = problem.evaluate(&candidate);
+            evaluations += 1;
+            // Accept on direction-correct improvement (Sphere is minimize).
+            let accept = candidate_eval.objectives[0] < current_eval.objectives[0];
+            if accept {
+                current = candidate;
+                current_eval = candidate_eval;
+            }
+        }
+
+        let best = Candidate::new(current.clone(), current_eval.clone());
+        let population = Population::new(vec![best.clone()]);
+        let pareto_front = vec![best.clone()];
+        OptimizationResult::new(
+            population,
+            pareto_front,
+            Some(best),
+            evaluations,
+            self.iterations,
+        )
+    }
+}
+
+struct Sphere1D;
+
+impl Problem for Sphere1D {
+    type Decision = Vec<f64>;
+
+    fn objectives(&self) -> ObjectiveSpace {
+        ObjectiveSpace::new(vec![Objective::minimize("f")])
+    }
+
+    fn evaluate(&self, x: &Vec<f64>) -> Evaluation {
+        Evaluation::new(vec![x[0] * x[0]])
+    }
+}
+
+fn main() {
+    let mut climber = HillClimber {
+        iterations: 500,
+        sigma: 0.5,
+        initializer: RealBounds::new(vec![(-5.0, 5.0)]),
+        seed: 11,
+    };
+
+    let result = climber.run(&Sphere1D);
+    let best = result.best.expect("single-objective always has a best");
+    println!("Hill-climber best f = {:.6}", best.evaluation.objectives[0]);
+    println!("Total evaluations: {}", result.evaluations);
+}
diff --git a/examples/random_search.rs b/examples/random_search.rs
new file mode 100644
index 0000000..1739e26
--- /dev/null
+++ b/examples/random_search.rs
@@ -0,0 +1,36 @@
+//! Run a 2D sphere problem under `RandomSearch`.
+//!
+//! Run with:
+//!
+//! ```bash
+//! cargo run --example random_search
+//! ```
+
+use heuropt::prelude::*;
+
+struct Sphere2D;
+
+impl Problem for Sphere2D {
+    type Decision = Vec<f64>;
+
+    fn objectives(&self) -> ObjectiveSpace {
+        ObjectiveSpace::new(vec![Objective::minimize("f")])
+    }
+
+    fn evaluate(&self, x: &Vec<f64>) -> Evaluation {
+        Evaluation::new(vec![x.iter().map(|v| v * v).sum()])
+    }
+}
+
+fn main() {
+    let initializer = RealBounds::new(vec![(-5.0, 5.0), (-5.0, 5.0)]);
+    let config = RandomSearchConfig { iterations: 500, batch_size: 1, seed: 7 };
+    let mut optimizer = RandomSearch::new(config, initializer);
+
+    let result = optimizer.run(&Sphere2D);
+
+    let best = result.best.expect("single-objective always has a best");
+    println!("Total evaluations: {}", result.evaluations);
+    println!("Best decision: {:?}", best.decision);
+    println!("Best f: {:.6}", best.evaluation.objectives[0]);
+}
diff --git a/examples/toy_nsga2.rs b/examples/toy_nsga2.rs
new file mode 100644
index 0000000..795ff33
--- /dev/null
+++ b/examples/toy_nsga2.rs
@@ -0,0 +1,42 @@
+//! Solve the Schaffer N.1 two-objective problem with NSGA-II.
+//!
+//! Run with:
+//!
+//! ```bash
+//! cargo run --example toy_nsga2
+//! ```
+
+use heuropt::prelude::*;
+
+struct SchafferN1;
+
+impl Problem for SchafferN1 {
+    type Decision = Vec<f64>;
+
+    fn objectives(&self) -> ObjectiveSpace {
+        ObjectiveSpace::new(vec![Objective::minimize("f1"), Objective::minimize("f2")])
+    }
+
+    fn evaluate(&self, x: &Vec<f64>) -> Evaluation {
+        let v = x[0];
+        Evaluation::new(vec![v * v, (v - 2.0).powi(2)])
+    }
+}
+
+fn main() {
+    let initializer = RealBounds::new(vec![(-5.0, 5.0)]);
+    let variation = GaussianMutation { sigma: 0.2 };
+    let config = Nsga2Config { population_size: 60, generations: 80, seed: 42 };
+    let mut optimizer = Nsga2::new(config, initializer, variation);
+
+    let result = optimizer.run(&SchafferN1);
+
+    println!("Final population: {}", result.population.len());
+    println!("Pareto front size: {}", result.pareto_front.len());
+    println!("Total evaluations: {}", result.evaluations);
+    println!("First few front points (f1, f2):");
+    for c in result.pareto_front.iter().take(8) {
+        let o = &c.evaluation.objectives;
+        println!("  ({:.4}, {:.4})", o[0], o[1]);
+    }
+}