chore: release v0.6.0

Add Study::enqueue() to push specific parameter configurations onto a
FIFO queue. The next call to ask()/create_trial() or the next iteration
in optimize() pops the front entry and injects it into the trial so
that suggest_param() returns the pre-filled value instead of sampling.

Parameters missing from an enqueued map fall back to normal sampling,
and once the queue is drained regular sampler-driven trials resume.

Cargo.toml

@@ -3,7 +3,7 @@ members = ["optimizer-derive"]
 
 [package]
 name = "optimizer"
-version = "0.5.1"
+version = "0.6.0"
 edition = "2024"
 rust-version = "1.88"
 license = "MIT"

src/error.rs

@@ -72,6 +72,10 @@ pub enum Error {
         got: usize,
     },
 
+    /// Returned when a trial is pruned (stopped early by the objective function).
+    #[error("trial was pruned")]
+    TrialPruned,
+
     /// Returned when an internal invariant is violated.
     #[error("internal error: {0}")]
     Internal(&'static str),
@@ -83,3 +87,33 @@ pub enum Error {
 }
 
 pub type Result<T> = core::result::Result<T, Error>;
+
+/// Convenience type for signalling a pruned trial from an objective function.
+///
+/// Implements `Into<Error>` so it can be used with `?` in objectives that
+/// return `Result<V, Error>`.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::{Error, TrialPruned};
+///
+/// fn objective_that_prunes() -> Result<f64, Error> {
+///     // ... some computation ...
+///     Err(TrialPruned)?
+/// }
+/// ```
+#[derive(Debug)]
+pub struct TrialPruned;
+
+impl core::fmt::Display for TrialPruned {
+    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
+        write!(f, "trial was pruned")
+    }
+}
+
+impl From<TrialPruned> for Error {
+    fn from(_: TrialPruned) -> Self {
+        Error::TrialPruned
+    }
+}

src/lib.rs

@@ -188,24 +188,29 @@ mod error;
 mod kde;
 mod param;
 pub mod parameter;
+pub mod pruner;
 pub mod sampler;
 mod study;
 mod trial;
 mod types;
 
-pub use error::{Error, Result};
+pub use error::{Error, Result, TrialPruned};
 #[cfg(feature = "derive")]
 pub use optimizer_derive::Categorical;
 pub use param::ParamValue;
 pub use parameter::{
     BoolParam, Categorical, CategoricalParam, EnumParam, FloatParam, IntParam, ParamId, Parameter,
 };
+pub use pruner::{
+    HyperbandPruner, MedianPruner, NopPruner, PatientPruner, PercentilePruner, Pruner,
+    SuccessiveHalvingPruner, ThresholdPruner, WilcoxonPruner,
+};
 pub use sampler::CompletedTrial;
 pub use sampler::grid::GridSearchSampler;
 pub use sampler::random::RandomSampler;
 pub use sampler::tpe::TpeSampler;
 pub use study::Study;
-pub use trial::Trial;
+pub use trial::{AttrValue, Trial};
 pub use types::{Direction, TrialState};
 
 /// Convenient wildcard import for the most common types.
@@ -217,16 +222,20 @@ pub mod prelude {
     #[cfg(feature = "derive")]
     pub use optimizer_derive::Categorical as DeriveCategory;
 
-    pub use crate::error::{Error, Result};
+    pub use crate::error::{Error, Result, TrialPruned};
     pub use crate::param::ParamValue;
     pub use crate::parameter::{
         BoolParam, Categorical, CategoricalParam, EnumParam, FloatParam, IntParam, Parameter,
     };
+    pub use crate::pruner::{
+        HyperbandPruner, MedianPruner, NopPruner, PatientPruner, PercentilePruner, Pruner,
+        SuccessiveHalvingPruner, ThresholdPruner,
+    };
     pub use crate::sampler::CompletedTrial;
     pub use crate::sampler::grid::GridSearchSampler;
     pub use crate::sampler::random::RandomSampler;
     pub use crate::sampler::tpe::TpeSampler;
     pub use crate::study::Study;
-    pub use crate::trial::Trial;
+    pub use crate::trial::{AttrValue, Trial};
     pub use crate::types::Direction;
 }

src/parameter.rs

@@ -21,6 +21,7 @@
 //! ```
 
 use core::fmt::Debug;
+use core::ops::RangeInclusive;
 use core::sync::atomic::{AtomicU64, Ordering};
 
 use crate::distribution::{
@@ -187,6 +188,12 @@ impl FloatParam {
     }
 }
 
+impl From<RangeInclusive<f64>> for FloatParam {
+    fn from(range: RangeInclusive<f64>) -> Self {
+        FloatParam::new(*range.start(), *range.end())
+    }
+}
+
 impl Parameter for FloatParam {
     type Value = f64;
 
@@ -306,6 +313,12 @@ impl IntParam {
     }
 }
 
+impl From<RangeInclusive<i64>> for IntParam {
+    fn from(range: RangeInclusive<i64>) -> Self {
+        IntParam::new(*range.start(), *range.end())
+    }
+}
+
 impl Parameter for IntParam {
     type Value = i64;
 
@@ -992,4 +1005,34 @@ mod tests {
         let cloned = param.clone();
         assert_eq!(param.id(), cloned.id());
     }
+
+    #[test]
+    fn float_param_from_range() {
+        let param = FloatParam::from(0.0..=1.0);
+        assert_eq!(
+            param.distribution(),
+            Distribution::Float(FloatDistribution {
+                low: 0.0,
+                high: 1.0,
+                log_scale: false,
+                step: None,
+            })
+        );
+        assert_eq!(param.label(), format!("{param:?}"));
+    }
+
+    #[test]
+    fn int_param_from_range() {
+        let param = IntParam::from(1..=10);
+        assert_eq!(
+            param.distribution(),
+            Distribution::Int(IntDistribution {
+                low: 1,
+                high: 10,
+                log_scale: false,
+                step: None,
+            })
+        );
+        assert_eq!(param.label(), format!("{param:?}"));
+    }
 }

src/pruner/hyperband.rs

@@ -0,0 +1,559 @@
+use core::sync::atomic::{AtomicU64, Ordering};
+use std::collections::HashMap;
+use std::sync::Mutex;
+
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+use crate::types::{Direction, TrialState};
+
+/// Hyperband pruner that manages multiple Successive Halving brackets.
+///
+/// Hyperband addresses SHA's sensitivity to the `min_resource` choice by
+/// running multiple brackets, each with a different tradeoff between the
+/// number of configurations and the starting budget:
+///
+/// - Bracket 0: many trials, very small starting budget (aggressive pruning)
+/// - Bracket 1: fewer trials, larger starting budget (moderate pruning)
+/// - ...
+/// - Bracket `s_max`: few trials, full budget (no pruning)
+///
+/// Trials are assigned to brackets in round-robin fashion. Each bracket
+/// runs SHA with its own `min_resource` and rung schedule.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::pruner::HyperbandPruner;
+///
+/// let pruner = HyperbandPruner::new()
+///     .min_resource(1)
+///     .max_resource(81)
+///     .reduction_factor(3)
+///     .direction(Direction::Minimize);
+/// ```
+pub struct HyperbandPruner {
+    min_resource: u64,
+    max_resource: u64,
+    reduction_factor: u64,
+    direction: Direction,
+    /// Tracks which bracket each trial belongs to.
+    trial_brackets: Mutex<HashMap<u64, usize>>,
+    /// Counter for round-robin bracket assignment.
+    next_bracket: AtomicU64,
+}
+
+impl HyperbandPruner {
+    /// Create a new `HyperbandPruner` with default parameters.
+    ///
+    /// Defaults: `min_resource=1`, `max_resource=81`, `reduction_factor=3`,
+    /// `direction=Minimize`.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            min_resource: 1,
+            max_resource: 81,
+            reduction_factor: 3,
+            direction: Direction::Minimize,
+            trial_brackets: Mutex::new(HashMap::new()),
+            next_bracket: AtomicU64::new(0),
+        }
+    }
+
+    /// Set the minimum resource (budget) per trial.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `r` is 0.
+    #[must_use]
+    pub fn min_resource(mut self, r: u64) -> Self {
+        assert!(r > 0, "min_resource must be > 0, got {r}");
+        self.min_resource = r;
+        self
+    }
+
+    /// Set the maximum resource (budget) per trial.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `r` is 0.
+    #[must_use]
+    pub fn max_resource(mut self, r: u64) -> Self {
+        assert!(r > 0, "max_resource must be > 0, got {r}");
+        self.max_resource = r;
+        self
+    }
+
+    /// Set the reduction factor (eta). At each rung, the top 1/eta trials survive.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `eta` is less than 2.
+    #[must_use]
+    pub fn reduction_factor(mut self, eta: u64) -> Self {
+        assert!(eta >= 2, "reduction_factor must be >= 2, got {eta}");
+        self.reduction_factor = eta;
+        self
+    }
+
+    /// Set the optimization direction.
+    #[must_use]
+    pub fn direction(mut self, d: Direction) -> Self {
+        self.direction = d;
+        self
+    }
+
+    /// Compute `s_max = floor(log(max_resource / min_resource) / log(eta))`.
+    #[allow(
+        clippy::cast_precision_loss,
+        clippy::cast_possible_truncation,
+        clippy::cast_sign_loss
+    )]
+    fn s_max(&self) -> u64 {
+        let eta = self.reduction_factor as f64;
+        let ratio = self.max_resource as f64 / self.min_resource as f64;
+        (ratio.ln() / eta.ln()).floor() as u64
+    }
+
+    /// Compute the rung steps for a given bracket `s`.
+    ///
+    /// For bracket `s`, the starting resource is `max_resource / eta^(s_max - s)`,
+    /// and rungs are spaced at powers of eta from there up to `max_resource`.
+    #[allow(
+        clippy::cast_precision_loss,
+        clippy::cast_possible_truncation,
+        clippy::cast_sign_loss
+    )]
+    fn rung_steps_for_bracket(&self, bracket: usize) -> Vec<u64> {
+        let s_max = self.s_max();
+        let eta = self.reduction_factor as f64;
+
+        // Starting resource for this bracket
+        let exponent = s_max.saturating_sub(bracket as u64);
+        let min_resource_bracket =
+            (self.max_resource as f64 / eta.powi(exponent as i32)).ceil() as u64;
+
+        let mut steps = Vec::new();
+        let mut rung: u32 = 0;
+        while let Some(power) = self.reduction_factor.checked_pow(rung) {
+            let step = min_resource_bracket.saturating_mul(power);
+            if step > self.max_resource {
+                break;
+            }
+            steps.push(step);
+            rung += 1;
+        }
+        steps
+    }
+
+    /// Assign a trial to a bracket (round-robin) and return the bracket index.
+    #[allow(clippy::cast_possible_truncation)]
+    fn assign_bracket(&self, trial_id: u64) -> usize {
+        let n_brackets = (self.s_max() + 1) as usize;
+        let mut map = self.trial_brackets.lock().expect("lock poisoned");
+        *map.entry(trial_id).or_insert_with(|| {
+            let idx = self.next_bracket.fetch_add(1, Ordering::Relaxed);
+            (idx as usize) % n_brackets
+        })
+    }
+}
+
+impl Default for HyperbandPruner {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[allow(clippy::cast_precision_loss)]
+impl Pruner for HyperbandPruner {
+    fn should_prune(
+        &self,
+        trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        let bracket = self.assign_bracket(trial_id);
+        let rungs = self.rung_steps_for_bracket(bracket);
+
+        // Find the highest rung step <= current step
+        let Some(&rung_step) = rungs.iter().rev().find(|&&r| r <= step) else {
+            return false;
+        };
+
+        // Never prune at the last rung (full budget)
+        if rung_step >= self.max_resource {
+            return false;
+        }
+
+        // Get the current trial's value at this rung step
+        let current_value =
+            if let Some(&(_, v)) = intermediate_values.iter().find(|(s, _)| *s == rung_step) {
+                v
+            } else if let Some(&(_, v)) = intermediate_values
+                .iter()
+                .rev()
+                .find(|(s, _)| *s <= rung_step)
+            {
+                v
+            } else {
+                return false;
+            };
+
+        self.is_pruned_at_rung(current_value, rung_step, bracket, completed_trials)
+    }
+}
+
+impl HyperbandPruner {
+    /// Determine whether a trial should be pruned at the given rung within its bracket.
+    ///
+    /// Only compares against other trials in the same bracket.
+    #[allow(
+        clippy::cast_precision_loss,
+        clippy::cast_possible_truncation,
+        clippy::cast_sign_loss
+    )]
+    fn is_pruned_at_rung(
+        &self,
+        current_value: f64,
+        rung_step: u64,
+        bracket: usize,
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        let eta = self.reduction_factor as usize;
+
+        // Collect values at this rung step from trials in the same bracket
+        let map = self.trial_brackets.lock().expect("lock poisoned");
+        let mut values_at_rung: Vec<f64> = completed_trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete || t.state == TrialState::Pruned)
+            .filter(|t| map.get(&t.id).copied() == Some(bracket))
+            .filter_map(|t| {
+                t.intermediate_values
+                    .iter()
+                    .find(|(s, _)| *s == rung_step)
+                    .map(|(_, v)| *v)
+            })
+            .collect();
+        drop(map);
+
+        // Need at least eta trials to make a meaningful comparison
+        if values_at_rung.len() < eta {
+            return false;
+        }
+
+        values_at_rung.push(current_value);
+
+        values_at_rung
+            .sort_unstable_by(|a, b| a.partial_cmp(b).unwrap_or(core::cmp::Ordering::Equal));
+        if self.direction == Direction::Maximize {
+            values_at_rung.reverse();
+        }
+
+        let n_keep = (values_at_rung.len() as f64 / eta as f64).ceil() as usize;
+        let threshold_idx = n_keep.max(1) - 1;
+        let threshold = values_at_rung[threshold_idx];
+
+        match self.direction {
+            Direction::Minimize => current_value > threshold,
+            Direction::Maximize => current_value < threshold,
+        }
+    }
+}
+
+#[cfg(test)]
+#[allow(clippy::cast_precision_loss)]
+mod tests {
+    use super::*;
+
+    fn make_trial(id: u64, values: &[(u64, f64)]) -> CompletedTrial {
+        use std::collections::HashMap;
+
+        use crate::parameter::ParamId;
+
+        CompletedTrial::with_intermediate_values(
+            id,
+            HashMap::<ParamId, crate::ParamValue>::new(),
+            HashMap::new(),
+            HashMap::new(),
+            0.0,
+            values.to_vec(),
+            HashMap::new(),
+        )
+    }
+
+    fn make_pruned_trial(id: u64, values: &[(u64, f64)]) -> CompletedTrial {
+        let mut t = make_trial(id, values);
+        t.state = TrialState::Pruned;
+        t
+    }
+
+    #[test]
+    fn s_max_default() {
+        let pruner = HyperbandPruner::new();
+        // s_max = floor(ln(81/1) / ln(3)) = floor(4.0) = 4
+        assert_eq!(pruner.s_max(), 4);
+    }
+
+    #[test]
+    fn s_max_custom() {
+        let pruner = HyperbandPruner::new()
+            .min_resource(1)
+            .max_resource(16)
+            .reduction_factor(2);
+        // s_max = floor(ln(16) / ln(2)) = floor(4.0) = 4
+        assert_eq!(pruner.s_max(), 4);
+    }
+
+    #[test]
+    fn bracket_count() {
+        let pruner = HyperbandPruner::new();
+        // s_max=4, so brackets 0..=4 → 5 brackets
+        assert_eq!(pruner.s_max() + 1, 5);
+    }
+
+    #[test]
+    fn rung_steps_bracket_0_default() {
+        let pruner = HyperbandPruner::new();
+        // Bracket 0: min_resource_bracket = ceil(81 / 3^4) = ceil(81/81) = 1
+        // Rungs: 1, 3, 9, 27, 81
+        assert_eq!(pruner.rung_steps_for_bracket(0), vec![1, 3, 9, 27, 81]);
+    }
+
+    #[test]
+    fn rung_steps_bracket_2_default() {
+        let pruner = HyperbandPruner::new();
+        // Bracket 2: min_resource_bracket = ceil(81 / 3^(4-2)) = ceil(81/9) = 9
+        // Rungs: 9, 27, 81
+        assert_eq!(pruner.rung_steps_for_bracket(2), vec![9, 27, 81]);
+    }
+
+    #[test]
+    fn rung_steps_bracket_4_default() {
+        let pruner = HyperbandPruner::new();
+        // Bracket 4 (s_max): min_resource_bracket = ceil(81 / 3^0) = 81
+        // Rungs: 81 only (no pruning, full budget)
+        assert_eq!(pruner.rung_steps_for_bracket(4), vec![81]);
+    }
+
+    #[test]
+    fn rung_steps_eta2() {
+        let pruner = HyperbandPruner::new()
+            .min_resource(1)
+            .max_resource(16)
+            .reduction_factor(2);
+        // s_max = 4
+        // Bracket 0: min=ceil(16/2^4)=1, rungs: 1,2,4,8,16
+        assert_eq!(pruner.rung_steps_for_bracket(0), vec![1, 2, 4, 8, 16]);
+        // Bracket 2: min=ceil(16/2^2)=4, rungs: 4,8,16
+        assert_eq!(pruner.rung_steps_for_bracket(2), vec![4, 8, 16]);
+        // Bracket 4: min=16, rungs: 16
+        assert_eq!(pruner.rung_steps_for_bracket(4), vec![16]);
+    }
+
+    #[test]
+    fn round_robin_bracket_assignment() {
+        let pruner = HyperbandPruner::new(); // 5 brackets (0..=4)
+        // Trials get assigned in round-robin: 0→0, 1→1, 2→2, 3→3, 4→4, 5→0, ...
+        assert_eq!(pruner.assign_bracket(100), 0);
+        assert_eq!(pruner.assign_bracket(101), 1);
+        assert_eq!(pruner.assign_bracket(102), 2);
+        assert_eq!(pruner.assign_bracket(103), 3);
+        assert_eq!(pruner.assign_bracket(104), 4);
+        assert_eq!(pruner.assign_bracket(105), 0); // wraps around
+
+        // Repeated calls for same trial return same bracket
+        assert_eq!(pruner.assign_bracket(100), 0);
+        assert_eq!(pruner.assign_bracket(103), 3);
+    }
+
+    #[test]
+    fn no_prune_before_first_rung() {
+        let pruner = HyperbandPruner::new().direction(Direction::Minimize);
+        // Assign trial 0 to bracket 0 (rungs: 1, 3, 9, 27, 81)
+        pruner.assign_bracket(0);
+
+        // Register completed trials in bracket 0
+        let mut completed = Vec::new();
+        for i in 1..=9 {
+            pruner.assign_bracket(i);
+            completed.push(make_trial(i, &[(1, i as f64)]));
+        }
+
+        // Trial at step 0 (before rung 1) → don't prune
+        assert!(!pruner.should_prune(0, 0, &[(0, 100.0)], &completed));
+    }
+
+    #[test]
+    fn no_prune_at_max_resource() {
+        let pruner = HyperbandPruner::new().direction(Direction::Minimize);
+
+        // Put all trials in bracket 0
+        let mut completed = Vec::new();
+        for i in 0..9 {
+            pruner.assign_bracket(i);
+            completed.push(make_trial(i, &[(81, (i + 1) as f64)]));
+        }
+
+        let trial_id = 9;
+        pruner.assign_bracket(trial_id);
+        // At max_resource (81), never prune
+        assert!(!pruner.should_prune(trial_id, 81, &[(81, 100.0)], &completed));
+    }
+
+    #[test]
+    fn prune_worst_in_bracket_minimize() {
+        let pruner = HyperbandPruner::new().direction(Direction::Minimize);
+
+        // Force all trials into bracket 0 by assigning sequentially
+        // With 5 brackets, trials 0,5,10,... go to bracket 0
+        let bracket_0_ids: Vec<u64> = (0..5).map(|i| i * 5).collect();
+        // Assign all 25 trial IDs to fill brackets
+        for i in 0..25 {
+            pruner.assign_bracket(i);
+        }
+
+        // Create 9 completed trials in bracket 0 at rung step=1
+        let completed: Vec<_> = bracket_0_ids
+            .iter()
+            .take(3)
+            .enumerate()
+            .map(|(idx, &id)| make_trial(id, &[(1, (idx + 1) as f64)]))
+            .collect();
+
+        // Trial 25 → bracket 0 (25 % 5 == 0)
+        let test_id = 25;
+        pruner.assign_bracket(test_id);
+        assert_eq!(pruner.assign_bracket(test_id), 0);
+
+        // 3 completed + 1 current = 4. eta=3. ceil(4/3)=2. Threshold = 2.0
+        // Value 2.0 → keep
+        assert!(!pruner.should_prune(test_id, 1, &[(1, 2.0)], &completed));
+        // Value 3.0 → prune
+        assert!(pruner.should_prune(test_id, 1, &[(1, 3.0)], &completed));
+    }
+
+    #[test]
+    fn prune_worst_in_bracket_maximize() {
+        let pruner = HyperbandPruner::new().direction(Direction::Maximize);
+
+        // Assign trials so they end up in bracket 0
+        for i in 0..25 {
+            pruner.assign_bracket(i);
+        }
+
+        let completed: Vec<_> = [0u64, 5, 10]
+            .iter()
+            .enumerate()
+            .map(|(idx, &id)| make_trial(id, &[(1, (idx + 1) as f64)]))
+            .collect();
+
+        let test_id = 25;
+        pruner.assign_bracket(test_id);
+
+        // For maximize, best = highest. Values: 1,2,3 + current
+        // Value 2.0 → keep (threshold = 2.0 when sorted desc: 3,2,current,1)
+        assert!(!pruner.should_prune(test_id, 1, &[(1, 2.0)], &completed));
+        // Value 1.0 → prune
+        assert!(pruner.should_prune(test_id, 1, &[(1, 0.5)], &completed));
+    }
+
+    #[test]
+    fn different_brackets_have_different_aggressiveness() {
+        let pruner = HyperbandPruner::new()
+            .min_resource(1)
+            .max_resource(81)
+            .reduction_factor(3)
+            .direction(Direction::Minimize);
+
+        let rungs_0 = pruner.rung_steps_for_bracket(0);
+        let rungs_2 = pruner.rung_steps_for_bracket(2);
+        let rungs_4 = pruner.rung_steps_for_bracket(4);
+
+        // Bracket 0 has the most rungs (most aggressive)
+        assert!(rungs_0.len() > rungs_2.len());
+        // Bracket 4 has just 1 rung (no pruning)
+        assert_eq!(rungs_4.len(), 1);
+        // Bracket 0 starts earliest
+        assert!(rungs_0[0] < rungs_2[0]);
+    }
+
+    #[test]
+    fn trials_in_different_brackets_independent() {
+        let pruner = HyperbandPruner::new().direction(Direction::Minimize);
+
+        // Assign trials: bracket 0 gets IDs 0,5,10,15,20
+        for i in 0..25 {
+            pruner.assign_bracket(i);
+        }
+
+        // Bracket 0 trials: bad values at rung step=1
+        let bracket_0_trials: Vec<_> = [0u64, 5, 10]
+            .iter()
+            .map(|&id| make_trial(id, &[(1, 100.0)]))
+            .collect();
+
+        // Bracket 1 trials: good values at rung step=1
+        let bracket_1_trials: Vec<_> = [1u64, 6, 11]
+            .iter()
+            .map(|&id| make_trial(id, &[(1, 1.0)]))
+            .collect();
+
+        let mut all_trials = bracket_0_trials;
+        all_trials.extend(bracket_1_trials);
+
+        // A new bracket-0 trial with value 50 should be compared against
+        // bracket-0 peers (100,100,100), not bracket-1 peers (1,1,1)
+        let test_id = 25; // bracket 0
+        pruner.assign_bracket(test_id);
+        // 3 peers at 100.0 + current at 50.0. ceil(4/3)=2. Sorted: 50,100,100,100. Threshold=100.0
+        // Value 50.0 < 100.0 → keep
+        assert!(!pruner.should_prune(test_id, 1, &[(1, 50.0)], &all_trials));
+    }
+
+    #[test]
+    fn includes_pruned_trials() {
+        let pruner = HyperbandPruner::new().direction(Direction::Minimize);
+
+        for i in 0..25 {
+            pruner.assign_bracket(i);
+        }
+
+        let completed = vec![
+            make_trial(0, &[(1, 1.0)]),
+            make_pruned_trial(5, &[(1, 8.0)]),
+            make_pruned_trial(10, &[(1, 9.0)]),
+        ];
+
+        let test_id = 25;
+        pruner.assign_bracket(test_id);
+
+        // Values: 1.0, 8.0, 9.0 + current. eta=3.
+        // Value 1.0 → keep
+        assert!(!pruner.should_prune(test_id, 1, &[(1, 1.0)], &completed));
+        // Value 5.0 → prune (sorted: 1,5,8,9 → keep ceil(4/3)=2 → threshold=5.0, 5.0 not > 5.0 → keep)
+        assert!(!pruner.should_prune(test_id, 1, &[(1, 5.0)], &completed));
+        // Value 6.0 → prune (sorted: 1,6,8,9 → threshold=6.0, 6.0 not > 6.0 → keep)
+        assert!(!pruner.should_prune(test_id, 1, &[(1, 6.0)], &completed));
+        // Value 9.5 → prune
+        assert!(pruner.should_prune(test_id, 1, &[(1, 9.5)], &completed));
+    }
+
+    #[test]
+    #[should_panic(expected = "min_resource must be > 0")]
+    fn rejects_zero_min_resource() {
+        let _ = HyperbandPruner::new().min_resource(0);
+    }
+
+    #[test]
+    #[should_panic(expected = "max_resource must be > 0")]
+    fn rejects_zero_max_resource() {
+        let _ = HyperbandPruner::new().max_resource(0);
+    }
+
+    #[test]
+    #[should_panic(expected = "reduction_factor must be >= 2")]
+    fn rejects_reduction_factor_one() {
+        let _ = HyperbandPruner::new().reduction_factor(1);
+    }
+}

src/pruner/median.rs

@@ -0,0 +1,127 @@
+use super::Pruner;
+use super::percentile::compute_percentile;
+use crate::sampler::CompletedTrial;
+use crate::types::{Direction, TrialState};
+
+/// Prune trials that are performing worse than the median of completed trials
+/// at the same step.
+///
+/// This is the most commonly used pruner. It compares the current trial's
+/// intermediate value at each step with the median of all completed trials'
+/// values at that same step.
+///
+/// Equivalent to `PercentilePruner::new(50.0, direction)`.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::pruner::MedianPruner;
+///
+/// // Prune trials worse than median when minimizing, after 5 warmup steps
+/// let pruner = MedianPruner::new(Direction::Minimize)
+///     .n_warmup_steps(5)
+///     .n_min_trials(3);
+/// ```
+pub struct MedianPruner {
+    /// The optimization direction.
+    direction: Direction,
+    /// Don't prune in the first N steps (let the trial warm up).
+    n_warmup_steps: u64,
+    /// Require at least N completed trials before pruning.
+    n_min_trials: usize,
+}
+
+impl MedianPruner {
+    /// Create a new `MedianPruner` for the given optimization direction.
+    ///
+    /// By default, `n_warmup_steps` is 0 and `n_min_trials` is 1.
+    #[must_use]
+    pub fn new(direction: Direction) -> Self {
+        Self {
+            direction,
+            n_warmup_steps: 0,
+            n_min_trials: 1,
+        }
+    }
+
+    /// Set the number of warmup steps. No pruning occurs before this step.
+    #[must_use]
+    pub fn n_warmup_steps(mut self, n: u64) -> Self {
+        self.n_warmup_steps = n;
+        self
+    }
+
+    /// Set the minimum number of completed trials required before pruning.
+    #[must_use]
+    pub fn n_min_trials(mut self, n: usize) -> Self {
+        self.n_min_trials = n;
+        self
+    }
+}
+
+impl Pruner for MedianPruner {
+    fn should_prune(
+        &self,
+        _trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        // 1. Don't prune during warmup
+        if step < self.n_warmup_steps {
+            return false;
+        }
+
+        // Get the current trial's latest value
+        let Some(&(_, current_value)) = intermediate_values.last() else {
+            return false;
+        };
+
+        // 2. Collect values at this step from completed (non-pruned) trials
+        let mut values_at_step: Vec<f64> = completed_trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .filter_map(|t| {
+                t.intermediate_values
+                    .iter()
+                    .find(|(s, _)| *s == step)
+                    .map(|(_, v)| *v)
+            })
+            .collect();
+
+        // 3. Not enough trials
+        if values_at_step.len() < self.n_min_trials {
+            return false;
+        }
+
+        // 4. Compute median (50th percentile)
+        let median = compute_percentile(&mut values_at_step, 50.0);
+
+        // 5. Compare against median based on direction
+        match self.direction {
+            Direction::Minimize => current_value > median,
+            Direction::Maximize => current_value < median,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn compute_median_odd() {
+        assert!((compute_percentile(&mut [3.0, 1.0, 2.0], 50.0) - 2.0).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn compute_median_even() {
+        assert!((compute_percentile(&mut [4.0, 1.0, 3.0, 2.0], 50.0) - 2.5).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn compute_median_single() {
+        assert!((compute_percentile(&mut [5.0], 50.0) - 5.0).abs() < f64::EPSILON);
+    }
+}

src/pruner/mod.rs

@@ -0,0 +1,74 @@
+//! Pruner trait and implementations for trial pruning.
+//!
+//! Pruners decide whether to stop (prune) a trial early based on its
+//! intermediate values compared to other trials. This is useful for
+//! discarding unpromising trials before they complete, saving compute.
+
+mod hyperband;
+mod median;
+mod nop;
+mod patient;
+pub(crate) mod percentile;
+mod successive_halving;
+mod threshold;
+mod wilcoxon;
+
+pub use hyperband::HyperbandPruner;
+pub use median::MedianPruner;
+pub use nop::NopPruner;
+pub use patient::PatientPruner;
+pub use percentile::PercentilePruner;
+pub use successive_halving::SuccessiveHalvingPruner;
+pub use threshold::ThresholdPruner;
+pub use wilcoxon::WilcoxonPruner;
+
+use crate::sampler::CompletedTrial;
+
+/// Trait for pluggable trial pruning strategies.
+///
+/// Pruners are consulted after each intermediate value is reported to
+/// decide whether the trial should be stopped early. The trait requires
+/// `Send + Sync` to support concurrent and async optimization.
+///
+/// # Implementing a custom pruner
+///
+/// ```
+/// use optimizer::pruner::Pruner;
+/// use optimizer::sampler::CompletedTrial;
+///
+/// struct MyPruner {
+///     threshold: f64,
+/// }
+///
+/// impl Pruner for MyPruner {
+///     fn should_prune(
+///         &self,
+///         _trial_id: u64,
+///         _step: u64,
+///         intermediate_values: &[(u64, f64)],
+///         _completed_trials: &[CompletedTrial],
+///     ) -> bool {
+///         // Prune if the latest value exceeds the threshold
+///         intermediate_values
+///             .last()
+///             .is_some_and(|&(_, v)| v > self.threshold)
+///     }
+/// }
+/// ```
+pub trait Pruner: Send + Sync {
+    /// Decide whether to prune a trial at the given step.
+    ///
+    /// # Arguments
+    ///
+    /// * `trial_id` - The current trial's ID.
+    /// * `step` - The step at which the intermediate value was reported.
+    /// * `intermediate_values` - All `(step, value)` pairs reported so far for this trial.
+    /// * `completed_trials` - History of all completed trials (for comparison).
+    fn should_prune(
+        &self,
+        trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool;
+}

src/pruner/nop.rs

@@ -0,0 +1,17 @@
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+
+/// A pruner that never prunes. This is the default when no pruner is configured.
+pub struct NopPruner;
+
+impl Pruner for NopPruner {
+    fn should_prune(
+        &self,
+        _trial_id: u64,
+        _step: u64,
+        _intermediate_values: &[(u64, f64)],
+        _completed_trials: &[CompletedTrial],
+    ) -> bool {
+        false
+    }
+}

src/pruner/patient.rs

@@ -0,0 +1,160 @@
+use std::collections::HashMap;
+use std::sync::Mutex;
+
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+
+/// Wraps another pruner and adds a patience window.
+///
+/// The inner pruner must recommend pruning for `patience` consecutive
+/// steps before this pruner actually prunes the trial. This is useful
+/// to prevent premature pruning when intermediate values are noisy.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::pruner::{PatientPruner, ThresholdPruner};
+///
+/// // Only prune after the threshold pruner recommends pruning 3 times in a row
+/// let inner = ThresholdPruner::new().upper(100.0);
+/// let pruner = PatientPruner::new(inner, 3);
+/// ```
+pub struct PatientPruner {
+    inner: Box<dyn Pruner>,
+    patience: u64,
+    /// Track consecutive prune recommendations per trial.
+    consecutive_counts: Mutex<HashMap<u64, u64>>,
+}
+
+impl PatientPruner {
+    /// Create a new `PatientPruner` wrapping the given inner pruner.
+    ///
+    /// The inner pruner must recommend pruning for `patience` consecutive
+    /// calls before this pruner returns `true`.
+    pub fn new(inner: impl Pruner + 'static, patience: u64) -> Self {
+        Self {
+            inner: Box::new(inner),
+            patience,
+            consecutive_counts: Mutex::new(HashMap::new()),
+        }
+    }
+}
+
+impl Pruner for PatientPruner {
+    fn should_prune(
+        &self,
+        trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        let inner_says_prune =
+            self.inner
+                .should_prune(trial_id, step, intermediate_values, completed_trials);
+        let mut counts = self.consecutive_counts.lock().expect("lock poisoned");
+        let count = counts.entry(trial_id).or_insert(0);
+        if inner_says_prune {
+            *count += 1;
+            *count >= self.patience
+        } else {
+            *count = 0;
+            false
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::pruner::ThresholdPruner;
+
+    /// A test pruner that always returns the given value.
+    struct ConstPruner(bool);
+
+    impl Pruner for ConstPruner {
+        fn should_prune(
+            &self,
+            _trial_id: u64,
+            _step: u64,
+            _intermediate_values: &[(u64, f64)],
+            _completed_trials: &[CompletedTrial],
+        ) -> bool {
+            self.0
+        }
+    }
+
+    /// A pruner that returns values from a sequence.
+    struct SequencePruner(Mutex<Vec<bool>>);
+
+    impl Pruner for SequencePruner {
+        fn should_prune(
+            &self,
+            _trial_id: u64,
+            _step: u64,
+            _intermediate_values: &[(u64, f64)],
+            _completed_trials: &[CompletedTrial],
+        ) -> bool {
+            self.0.lock().expect("lock poisoned").remove(0)
+        }
+    }
+
+    fn call(pruner: &PatientPruner, trial_id: u64, step: u64) -> bool {
+        pruner.should_prune(trial_id, step, &[(step, 0.0)], &[])
+    }
+
+    #[test]
+    fn patience_1_behaves_like_inner() {
+        let pruner = PatientPruner::new(ConstPruner(true), 1);
+        assert!(call(&pruner, 0, 0));
+        assert!(call(&pruner, 0, 1));
+
+        let pruner = PatientPruner::new(ConstPruner(false), 1);
+        assert!(!call(&pruner, 0, 0));
+        assert!(!call(&pruner, 0, 1));
+    }
+
+    #[test]
+    fn patience_3_requires_consecutive_recommendations() {
+        let pruner = PatientPruner::new(ConstPruner(true), 3);
+        assert!(!call(&pruner, 0, 0)); // count=1
+        assert!(!call(&pruner, 0, 1)); // count=2
+        assert!(call(&pruner, 0, 2)); // count=3 → prune
+    }
+
+    #[test]
+    fn counter_resets_on_no_prune() {
+        // Sequence: prune, prune, no-prune, prune, prune, prune
+        let seq = vec![true, true, false, true, true, true];
+        let pruner = PatientPruner::new(SequencePruner(Mutex::new(seq)), 3);
+
+        assert!(!call(&pruner, 0, 0)); // count=1
+        assert!(!call(&pruner, 0, 1)); // count=2
+        assert!(!call(&pruner, 0, 2)); // reset → count=0
+        assert!(!call(&pruner, 0, 3)); // count=1
+        assert!(!call(&pruner, 0, 4)); // count=2
+        assert!(call(&pruner, 0, 5)); // count=3 → prune
+    }
+
+    #[test]
+    fn independent_per_trial() {
+        let pruner = PatientPruner::new(ConstPruner(true), 2);
+        assert!(!call(&pruner, 0, 0)); // trial 0: count=1
+        assert!(!call(&pruner, 1, 0)); // trial 1: count=1
+        assert!(call(&pruner, 0, 1)); // trial 0: count=2 → prune
+        assert!(!call(&pruner, 2, 0)); // trial 2: count=1
+        assert!(call(&pruner, 1, 1)); // trial 1: count=2 → prune
+    }
+
+    #[test]
+    fn works_with_threshold_pruner() {
+        let inner = ThresholdPruner::new().upper(10.0);
+        let pruner = PatientPruner::new(inner, 2);
+
+        // Value below threshold → inner says no
+        assert!(!pruner.should_prune(0, 0, &[(0, 5.0)], &[]));
+        // Value above threshold → inner says yes, count=1
+        assert!(!pruner.should_prune(0, 1, &[(0, 5.0), (1, 15.0)], &[]));
+        // Value above threshold again → count=2 → prune
+        assert!(pruner.should_prune(0, 2, &[(0, 5.0), (1, 15.0), (2, 20.0)], &[]));
+    }
+}

src/pruner/percentile.rs

@@ -0,0 +1,294 @@
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+use crate::types::{Direction, TrialState};
+
+/// Prune trials that are not in the top `percentile`% of completed trials
+/// at the same training step.
+///
+/// `PercentilePruner::new(50.0, direction)` is equivalent to `MedianPruner`.
+/// `PercentilePruner::new(25.0, direction)` keeps only the top 25% of trials.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::pruner::PercentilePruner;
+///
+/// // Keep only the top 25% of trials (aggressive pruning)
+/// let pruner = PercentilePruner::new(25.0, Direction::Minimize)
+///     .n_warmup_steps(5)
+///     .n_min_trials(3);
+/// ```
+pub struct PercentilePruner {
+    /// Keep trials in the top `percentile`%. Range: (0.0, 100.0).
+    percentile: f64,
+    /// Don't prune in the first N steps (let the trial warm up).
+    n_warmup_steps: u64,
+    /// Require at least N completed trials before pruning.
+    n_min_trials: usize,
+    /// The optimization direction.
+    direction: Direction,
+}
+
+impl PercentilePruner {
+    /// Create a new `PercentilePruner` for the given percentile and direction.
+    ///
+    /// The `percentile` value must be in `(0.0, 100.0)`.
+    /// A percentile of 50.0 is equivalent to median pruning.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `percentile` is not in `(0.0, 100.0)`.
+    #[must_use]
+    pub fn new(percentile: f64, direction: Direction) -> Self {
+        assert!(
+            percentile > 0.0 && percentile < 100.0,
+            "percentile must be in (0.0, 100.0), got {percentile}"
+        );
+        Self {
+            percentile,
+            n_warmup_steps: 0,
+            n_min_trials: 1,
+            direction,
+        }
+    }
+
+    /// Set the number of warmup steps. No pruning occurs before this step.
+    #[must_use]
+    pub fn n_warmup_steps(mut self, n: u64) -> Self {
+        self.n_warmup_steps = n;
+        self
+    }
+
+    /// Set the minimum number of completed trials required before pruning.
+    #[must_use]
+    pub fn n_min_trials(mut self, n: usize) -> Self {
+        self.n_min_trials = n;
+        self
+    }
+}
+
+impl Pruner for PercentilePruner {
+    fn should_prune(
+        &self,
+        _trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        // 1. Don't prune during warmup
+        if step < self.n_warmup_steps {
+            return false;
+        }
+
+        // Get the current trial's latest value
+        let Some(&(_, current_value)) = intermediate_values.last() else {
+            return false;
+        };
+
+        // 2. Collect values at this step from completed (non-pruned) trials
+        let mut values_at_step: Vec<f64> = completed_trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .filter_map(|t| {
+                t.intermediate_values
+                    .iter()
+                    .find(|(s, _)| *s == step)
+                    .map(|(_, v)| *v)
+            })
+            .collect();
+
+        // 3. Not enough trials
+        if values_at_step.len() < self.n_min_trials {
+            return false;
+        }
+
+        // 4. Compute percentile threshold
+        let threshold = compute_percentile(&mut values_at_step, self.percentile);
+
+        // 5. Compare against threshold based on direction
+        match self.direction {
+            Direction::Minimize => current_value > threshold,
+            Direction::Maximize => current_value < threshold,
+        }
+    }
+}
+
+/// Compute the given percentile of a non-empty slice. Sorts the slice in place.
+///
+/// Uses linear interpolation between the two nearest ranks.
+#[allow(
+    clippy::cast_precision_loss,
+    clippy::cast_possible_truncation,
+    clippy::cast_sign_loss
+)]
+pub(crate) fn compute_percentile(values: &mut [f64], percentile: f64) -> f64 {
+    values.sort_unstable_by(|a, b| a.partial_cmp(b).unwrap_or(core::cmp::Ordering::Equal));
+    let len = values.len();
+    if len == 1 {
+        return values[0];
+    }
+    // Rank in [0, len-1] range
+    let rank = percentile / 100.0 * (len - 1) as f64;
+    let lower = rank.floor() as usize;
+    let upper = rank.ceil() as usize;
+    if lower == upper {
+        values[lower]
+    } else {
+        let frac = rank - lower as f64;
+        values[lower] * (1.0 - frac) + values[upper] * frac
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn compute_percentile_median_odd() {
+        // Percentile 50 on odd-length slice = median
+        let val = compute_percentile(&mut [3.0, 1.0, 2.0], 50.0);
+        assert!((val - 2.0).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn compute_percentile_median_even() {
+        // Percentile 50 on even-length slice = median (interpolated)
+        let val = compute_percentile(&mut [4.0, 1.0, 3.0, 2.0], 50.0);
+        assert!((val - 2.5).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn compute_percentile_25() {
+        // [1.0, 2.0, 3.0, 4.0], rank = 0.25 * 3 = 0.75
+        // interpolate: 1.0 * 0.25 + 2.0 * 0.75 = 1.75
+        let val = compute_percentile(&mut [4.0, 1.0, 3.0, 2.0], 25.0);
+        assert!((val - 1.75).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn compute_percentile_75() {
+        // [1.0, 2.0, 3.0, 4.0], rank = 0.75 * 3 = 2.25
+        // interpolate: 3.0 * 0.75 + 4.0 * 0.25 = 3.25
+        let val = compute_percentile(&mut [4.0, 1.0, 3.0, 2.0], 75.0);
+        assert!((val - 3.25).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn compute_percentile_single() {
+        let val = compute_percentile(&mut [5.0], 50.0);
+        assert!((val - 5.0).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    #[should_panic(expected = "percentile must be in (0.0, 100.0)")]
+    fn new_rejects_zero() {
+        let _ = PercentilePruner::new(0.0, Direction::Minimize);
+    }
+
+    #[test]
+    #[should_panic(expected = "percentile must be in (0.0, 100.0)")]
+    fn new_rejects_hundred() {
+        let _ = PercentilePruner::new(100.0, Direction::Minimize);
+    }
+
+    fn make_completed_trial(id: u64, values: &[(u64, f64)]) -> CompletedTrial {
+        use std::collections::HashMap;
+
+        use crate::parameter::ParamId;
+
+        CompletedTrial::with_intermediate_values(
+            id,
+            HashMap::<ParamId, crate::ParamValue>::new(),
+            HashMap::new(),
+            HashMap::new(),
+            0.0,
+            values.to_vec(),
+            HashMap::new(),
+        )
+    }
+
+    #[test]
+    fn percentile_50_matches_median_behavior() {
+        let pruner = PercentilePruner::new(50.0, Direction::Minimize);
+        let completed = vec![
+            make_completed_trial(0, &[(0, 1.0), (1, 2.0)]),
+            make_completed_trial(1, &[(0, 3.0), (1, 4.0)]),
+            make_completed_trial(2, &[(0, 5.0), (1, 6.0)]),
+        ];
+        // Median at step 1 is 4.0
+        // Value 5.0 > 4.0 → prune
+        assert!(pruner.should_prune(3, 1, &[(0, 3.0), (1, 5.0)], &completed));
+        // Value 3.0 < 4.0 → keep
+        assert!(!pruner.should_prune(3, 1, &[(0, 3.0), (1, 3.0)], &completed));
+    }
+
+    #[test]
+    fn percentile_25_is_more_aggressive() {
+        let pruner_25 = PercentilePruner::new(25.0, Direction::Minimize);
+        let pruner_75 = PercentilePruner::new(75.0, Direction::Minimize);
+        let completed = vec![
+            make_completed_trial(0, &[(0, 1.0)]),
+            make_completed_trial(1, &[(0, 2.0)]),
+            make_completed_trial(2, &[(0, 3.0)]),
+            make_completed_trial(3, &[(0, 4.0)]),
+        ];
+        // 25th percentile at step 0: 1.75
+        // 75th percentile at step 0: 3.25
+        // Value 2.5: above 25th (prune), below 75th (keep)
+        assert!(pruner_25.should_prune(4, 0, &[(0, 2.5)], &completed));
+        assert!(!pruner_75.should_prune(4, 0, &[(0, 2.5)], &completed));
+    }
+
+    #[test]
+    fn warmup_prevents_pruning() {
+        let pruner = PercentilePruner::new(50.0, Direction::Minimize).n_warmup_steps(5);
+        let completed = vec![make_completed_trial(0, &[(0, 1.0)])];
+        // Step 3 < warmup 5 → no prune even with bad value
+        assert!(!pruner.should_prune(1, 3, &[(3, 100.0)], &completed));
+    }
+
+    #[test]
+    fn n_min_trials_prevents_pruning() {
+        let pruner = PercentilePruner::new(50.0, Direction::Minimize).n_min_trials(5);
+        let completed = vec![
+            make_completed_trial(0, &[(0, 1.0)]),
+            make_completed_trial(1, &[(0, 2.0)]),
+        ];
+        // Only 2 trials, need 5 → no prune
+        assert!(!pruner.should_prune(2, 0, &[(0, 100.0)], &completed));
+    }
+
+    #[test]
+    fn maximize_direction() {
+        let pruner = PercentilePruner::new(50.0, Direction::Maximize);
+        let completed = vec![
+            make_completed_trial(0, &[(0, 1.0)]),
+            make_completed_trial(1, &[(0, 3.0)]),
+            make_completed_trial(2, &[(0, 5.0)]),
+        ];
+        // Median at step 0 is 3.0
+        // Value 2.0 < 3.0 → prune (maximize wants higher)
+        assert!(pruner.should_prune(3, 0, &[(0, 2.0)], &completed));
+        // Value 4.0 > 3.0 → keep
+        assert!(!pruner.should_prune(3, 0, &[(0, 4.0)], &completed));
+    }
+
+    #[test]
+    fn near_boundary_percentiles() {
+        let pruner_low = PercentilePruner::new(1.0, Direction::Minimize);
+        let pruner_high = PercentilePruner::new(99.0, Direction::Minimize);
+        let completed = vec![
+            make_completed_trial(0, &[(0, 1.0)]),
+            make_completed_trial(1, &[(0, 2.0)]),
+            make_completed_trial(2, &[(0, 3.0)]),
+            make_completed_trial(3, &[(0, 100.0)]),
+        ];
+        // Percentile 1 is very aggressive (threshold near 1.0)
+        // Value 1.5 should be pruned
+        assert!(pruner_low.should_prune(4, 0, &[(0, 1.5)], &completed));
+        // Percentile 99 is very lenient (threshold near 100.0)
+        // Value 50.0 should not be pruned
+        assert!(!pruner_high.should_prune(4, 0, &[(0, 50.0)], &completed));
+    }
+}

src/pruner/successive_halving.rs

@@ -0,0 +1,466 @@
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+use crate::types::{Direction, TrialState};
+
+/// Successive Halving pruner based on the SHA algorithm.
+///
+/// Trials are evaluated at exponentially-spaced "rungs". At each rung,
+/// only the top 1/eta fraction of trials survive to the next rung.
+///
+/// For example, with `min_resource=1`, `max_resource=81`, `reduction_factor=3`:
+/// - Rung 0: evaluate at step 1, keep top 1/3
+/// - Rung 1: evaluate at step 3, keep top 1/3
+/// - Rung 2: evaluate at step 9, keep top 1/3
+/// - Rung 3: evaluate at step 27, keep top 1/3
+/// - Rung 4: evaluate at step 81 (full budget)
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::pruner::SuccessiveHalvingPruner;
+///
+/// let pruner = SuccessiveHalvingPruner::new()
+///     .min_resource(1)
+///     .max_resource(81)
+///     .reduction_factor(3)
+///     .direction(Direction::Minimize);
+/// ```
+pub struct SuccessiveHalvingPruner {
+    min_resource: u64,
+    max_resource: u64,
+    reduction_factor: u64,
+    min_early_stopping_rate: u64,
+    direction: Direction,
+}
+
+impl SuccessiveHalvingPruner {
+    /// Create a new `SuccessiveHalvingPruner` with default parameters.
+    ///
+    /// Defaults: `min_resource=1`, `max_resource=81`, `reduction_factor=3`,
+    /// `min_early_stopping_rate=0`, `direction=Minimize`.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            min_resource: 1,
+            max_resource: 81,
+            reduction_factor: 3,
+            min_early_stopping_rate: 0,
+            direction: Direction::Minimize,
+        }
+    }
+
+    /// Set the minimum resource (budget) per trial.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `r` is 0.
+    #[must_use]
+    pub fn min_resource(mut self, r: u64) -> Self {
+        assert!(r > 0, "min_resource must be > 0, got {r}");
+        self.min_resource = r;
+        self
+    }
+
+    /// Set the maximum resource (budget) per trial.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `r` is 0.
+    #[must_use]
+    pub fn max_resource(mut self, r: u64) -> Self {
+        assert!(r > 0, "max_resource must be > 0, got {r}");
+        self.max_resource = r;
+        self
+    }
+
+    /// Set the reduction factor (eta). At each rung, the top 1/eta trials survive.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `eta` is less than 2.
+    #[must_use]
+    pub fn reduction_factor(mut self, eta: u64) -> Self {
+        assert!(eta >= 2, "reduction_factor must be >= 2, got {eta}");
+        self.reduction_factor = eta;
+        self
+    }
+
+    /// Set the minimum early stopping rate. Skips the first N rungs.
+    #[must_use]
+    pub fn min_early_stopping_rate(mut self, n: u64) -> Self {
+        self.min_early_stopping_rate = n;
+        self
+    }
+
+    /// Set the optimization direction.
+    #[must_use]
+    pub fn direction(mut self, d: Direction) -> Self {
+        self.direction = d;
+        self
+    }
+
+    /// Compute the rung steps: `[min_resource * eta^(s), ...]` up to `max_resource`,
+    /// skipping the first `min_early_stopping_rate` rungs.
+    fn rung_steps(&self) -> Vec<u64> {
+        let eta = self.reduction_factor;
+        let mut steps = Vec::new();
+        let mut rung: u32 = 0;
+        while let Some(power) = eta.checked_pow(rung) {
+            let step = self.min_resource.saturating_mul(power);
+            if step > self.max_resource {
+                break;
+            }
+            if u64::from(rung) >= self.min_early_stopping_rate {
+                steps.push(step);
+            }
+            rung += 1;
+        }
+        steps
+    }
+}
+
+impl Default for SuccessiveHalvingPruner {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[allow(clippy::cast_precision_loss)]
+impl Pruner for SuccessiveHalvingPruner {
+    fn should_prune(
+        &self,
+        _trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        let rungs = self.rung_steps();
+
+        // Find the highest rung step <= current step
+        let Some(&rung_step) = rungs.iter().rev().find(|&&r| r <= step) else {
+            // No rung matches (before the first rung) → don't prune
+            return false;
+        };
+
+        // If this is the last rung (full budget), don't prune
+        if rung_step >= self.max_resource {
+            return false;
+        }
+
+        // Get the current trial's value at this rung step
+        let Some(&(_, current_value)) = intermediate_values.iter().find(|(s, _)| *s == rung_step)
+        else {
+            // Trial hasn't reported a value at this exact rung step.
+            // Use the latest intermediate value at or before the rung step instead.
+            let Some(&(_, current_value)) = intermediate_values
+                .iter()
+                .rev()
+                .find(|(s, _)| *s <= rung_step)
+            else {
+                return false;
+            };
+            return self.is_pruned_at_rung(current_value, rung_step, completed_trials);
+        };
+
+        self.is_pruned_at_rung(current_value, rung_step, completed_trials)
+    }
+}
+
+impl SuccessiveHalvingPruner {
+    /// Determine whether a trial with `current_value` should be pruned at the given rung.
+    #[allow(
+        clippy::cast_precision_loss,
+        clippy::cast_possible_truncation,
+        clippy::cast_sign_loss
+    )]
+    fn is_pruned_at_rung(
+        &self,
+        current_value: f64,
+        rung_step: u64,
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        let eta = self.reduction_factor as usize;
+
+        // Collect values at this rung step from all trials that reached it
+        let mut values_at_rung: Vec<f64> = completed_trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete || t.state == TrialState::Pruned)
+            .filter_map(|t| {
+                t.intermediate_values
+                    .iter()
+                    .find(|(s, _)| *s == rung_step)
+                    .map(|(_, v)| *v)
+            })
+            .collect();
+
+        // Need at least eta trials to make a meaningful comparison
+        // (with fewer trials, we can't determine the top 1/eta fraction)
+        if values_at_rung.len() < eta {
+            return false;
+        }
+
+        // Include the current trial's value for ranking
+        values_at_rung.push(current_value);
+
+        // Sort based on direction: best values first
+        values_at_rung
+            .sort_unstable_by(|a, b| a.partial_cmp(b).unwrap_or(core::cmp::Ordering::Equal));
+        if self.direction == Direction::Maximize {
+            values_at_rung.reverse();
+        }
+
+        // Keep top 1/eta fraction
+        let n_keep = (values_at_rung.len() as f64 / eta as f64).ceil() as usize;
+        let threshold_idx = n_keep.max(1) - 1;
+        let threshold = values_at_rung[threshold_idx];
+
+        // Prune if current value is worse than the threshold
+        match self.direction {
+            Direction::Minimize => current_value > threshold,
+            Direction::Maximize => current_value < threshold,
+        }
+    }
+}
+
+#[cfg(test)]
+#[allow(clippy::cast_precision_loss)]
+mod tests {
+    use super::*;
+
+    fn make_trial(id: u64, values: &[(u64, f64)]) -> CompletedTrial {
+        use std::collections::HashMap;
+
+        use crate::parameter::ParamId;
+
+        CompletedTrial::with_intermediate_values(
+            id,
+            HashMap::<ParamId, crate::ParamValue>::new(),
+            HashMap::new(),
+            HashMap::new(),
+            0.0,
+            values.to_vec(),
+            HashMap::new(),
+        )
+    }
+
+    fn make_pruned_trial(id: u64, values: &[(u64, f64)]) -> CompletedTrial {
+        let mut t = make_trial(id, values);
+        t.state = TrialState::Pruned;
+        t
+    }
+
+    #[test]
+    fn rung_steps_default() {
+        let pruner = SuccessiveHalvingPruner::new();
+        let rungs = pruner.rung_steps();
+        // min=1, max=81, eta=3 → 1, 3, 9, 27, 81
+        assert_eq!(rungs, vec![1, 3, 9, 27, 81]);
+    }
+
+    #[test]
+    fn rung_steps_custom() {
+        let pruner = SuccessiveHalvingPruner::new()
+            .min_resource(2)
+            .max_resource(32)
+            .reduction_factor(2);
+        let rungs = pruner.rung_steps();
+        // 2, 4, 8, 16, 32
+        assert_eq!(rungs, vec![2, 4, 8, 16, 32]);
+    }
+
+    #[test]
+    fn rung_steps_with_early_stopping_rate() {
+        let pruner = SuccessiveHalvingPruner::new().min_early_stopping_rate(2);
+        let rungs = pruner.rung_steps();
+        // Skip rung 0 (step=1) and rung 1 (step=3), keep rung 2+ (9, 27, 81)
+        assert_eq!(rungs, vec![9, 27, 81]);
+    }
+
+    #[test]
+    fn no_prune_before_first_rung() {
+        let pruner = SuccessiveHalvingPruner::new()
+            .min_resource(10)
+            .max_resource(100)
+            .reduction_factor(3);
+        let completed = vec![
+            make_trial(0, &[(5, 1.0)]),
+            make_trial(1, &[(5, 2.0)]),
+            make_trial(2, &[(5, 3.0)]),
+        ];
+        // Step 5 is before the first rung (10)
+        assert!(!pruner.should_prune(3, 5, &[(5, 100.0)], &completed));
+    }
+
+    #[test]
+    fn no_prune_with_single_trial() {
+        let pruner = SuccessiveHalvingPruner::new();
+        let completed = vec![make_trial(0, &[(1, 5.0)]), make_trial(1, &[(1, 3.0)])];
+        // Only 2 completed trials at rung + 1 current = 3 total, threshold = ceil(3/3) = 1
+        // With eta=3, we need at least 3 completed trials
+        assert!(!pruner.should_prune(2, 1, &[(1, 10.0)], &completed));
+    }
+
+    #[test]
+    fn prune_worst_trials_at_rung() {
+        let pruner = SuccessiveHalvingPruner::new().direction(Direction::Minimize);
+
+        // 9 completed trials at rung step=1, with values 1..=9
+        let completed: Vec<_> = (0..9)
+            .map(|i| make_trial(i, &[(1, (i + 1) as f64)]))
+            .collect();
+
+        // With eta=3, keep top 1/3. 10 total values → ceil(10/3) = 4 kept
+        // Best 4 values: 1, 2, 3, 4. Threshold = 4.0
+        // Value 3.0 → keep (in top 1/3)
+        assert!(!pruner.should_prune(9, 1, &[(1, 3.0)], &completed));
+        // Value 5.0 → prune (not in top 1/3)
+        assert!(pruner.should_prune(9, 1, &[(1, 5.0)], &completed));
+    }
+
+    #[test]
+    fn top_fraction_survives() {
+        let pruner = SuccessiveHalvingPruner::new().direction(Direction::Minimize);
+
+        // 6 completed trials at step=1
+        let completed: Vec<_> = (0..6)
+            .map(|i| make_trial(i, &[(1, (i + 1) as f64)]))
+            .collect();
+
+        // 7 total (6 + current). ceil(7/3) = 3 keep. Threshold = 3.0
+        // Value 2.0 → keep
+        assert!(!pruner.should_prune(6, 1, &[(1, 2.0)], &completed));
+        // Value 3.0 → keep (at threshold)
+        assert!(!pruner.should_prune(6, 1, &[(1, 3.0)], &completed));
+        // Value 4.0 → prune
+        assert!(pruner.should_prune(6, 1, &[(1, 4.0)], &completed));
+    }
+
+    #[test]
+    fn maximize_direction() {
+        let pruner = SuccessiveHalvingPruner::new().direction(Direction::Maximize);
+
+        let completed: Vec<_> = (0..6)
+            .map(|i| make_trial(i, &[(1, (i + 1) as f64)]))
+            .collect();
+
+        // 7 total. For maximize, best = highest. ceil(7/3)=3. Top 3: 6,5,4. Threshold=4.0
+        // Value 5.0 → keep
+        assert!(!pruner.should_prune(6, 1, &[(1, 5.0)], &completed));
+        // Value 4.0 → keep (at threshold)
+        assert!(!pruner.should_prune(6, 1, &[(1, 4.0)], &completed));
+        // Value 3.0 → prune
+        assert!(pruner.should_prune(6, 1, &[(1, 3.0)], &completed));
+    }
+
+    #[test]
+    fn reduction_factor_2() {
+        let pruner = SuccessiveHalvingPruner::new()
+            .reduction_factor(2)
+            .min_resource(1)
+            .max_resource(16)
+            .direction(Direction::Minimize);
+
+        // Rungs: 1, 2, 4, 8, 16
+        assert_eq!(pruner.rung_steps(), vec![1, 2, 4, 8, 16]);
+
+        // 4 completed trials at rung step=1
+        let completed: Vec<_> = (0..4)
+            .map(|i| make_trial(i, &[(1, (i + 1) as f64)]))
+            .collect();
+
+        // With eta=2, 5 total. ceil(5/2) = 3 keep. Threshold = 3.0
+        // Value 3.0 → keep
+        assert!(!pruner.should_prune(4, 1, &[(1, 3.0)], &completed));
+        // Value 4.0 → prune
+        assert!(pruner.should_prune(4, 1, &[(1, 4.0)], &completed));
+    }
+
+    #[test]
+    fn reduction_factor_4() {
+        let pruner = SuccessiveHalvingPruner::new()
+            .reduction_factor(4)
+            .min_resource(1)
+            .max_resource(64)
+            .direction(Direction::Minimize);
+
+        // Rungs: 1, 4, 16, 64
+        assert_eq!(pruner.rung_steps(), vec![1, 4, 16, 64]);
+
+        // 12 completed trials at rung step=1
+        let completed: Vec<_> = (0..12)
+            .map(|i| make_trial(i, &[(1, (i + 1) as f64)]))
+            .collect();
+
+        // With eta=4, 13 total. ceil(13/4) = 4 keep. Threshold = 4.0
+        // Value 4.0 → keep
+        assert!(!pruner.should_prune(12, 1, &[(1, 4.0)], &completed));
+        // Value 5.0 → prune
+        assert!(pruner.should_prune(12, 1, &[(1, 5.0)], &completed));
+    }
+
+    #[test]
+    fn non_contiguous_steps() {
+        let pruner = SuccessiveHalvingPruner::new().direction(Direction::Minimize);
+
+        // Trials reporting at rung step=3 (not step=1)
+        let completed: Vec<_> = (0..6)
+            .map(|i| make_trial(i, &[(3, (i + 1) as f64)]))
+            .collect();
+
+        // Current trial reports at step 5 (between rung 3 and rung 9)
+        // Highest rung <= 5 is 3. Use value at rung step 3.
+        // Trial has value at step 3 → use it
+        assert!(!pruner.should_prune(6, 5, &[(3, 2.0)], &completed));
+        assert!(pruner.should_prune(6, 5, &[(3, 5.0)], &completed));
+    }
+
+    #[test]
+    fn no_prune_at_max_resource() {
+        let pruner = SuccessiveHalvingPruner::new();
+        let completed: Vec<_> = (0..9)
+            .map(|i| make_trial(i, &[(81, (i + 1) as f64)]))
+            .collect();
+
+        // At the max resource rung, never prune (trial should complete)
+        assert!(!pruner.should_prune(9, 81, &[(81, 100.0)], &completed));
+    }
+
+    #[test]
+    fn includes_pruned_trials_in_comparison() {
+        let pruner = SuccessiveHalvingPruner::new().direction(Direction::Minimize);
+
+        // Mix of completed and pruned trials at rung step=1
+        let completed = vec![
+            make_trial(0, &[(1, 1.0)]),
+            make_trial(1, &[(1, 2.0)]),
+            make_pruned_trial(2, &[(1, 8.0)]),
+            make_pruned_trial(3, &[(1, 9.0)]),
+            make_pruned_trial(4, &[(1, 10.0)]),
+        ];
+
+        // 6 total. ceil(6/3) = 2 keep. Threshold = 2.0
+        // Value 2.0 → keep
+        assert!(!pruner.should_prune(5, 1, &[(1, 2.0)], &completed));
+        // Value 3.0 → prune
+        assert!(pruner.should_prune(5, 1, &[(1, 3.0)], &completed));
+    }
+
+    #[test]
+    #[should_panic(expected = "min_resource must be > 0")]
+    fn rejects_zero_min_resource() {
+        let _ = SuccessiveHalvingPruner::new().min_resource(0);
+    }
+
+    #[test]
+    #[should_panic(expected = "max_resource must be > 0")]
+    fn rejects_zero_max_resource() {
+        let _ = SuccessiveHalvingPruner::new().max_resource(0);
+    }
+
+    #[test]
+    #[should_panic(expected = "reduction_factor must be >= 2")]
+    fn rejects_reduction_factor_one() {
+        let _ = SuccessiveHalvingPruner::new().reduction_factor(1);
+    }
+}

src/pruner/threshold.rs

@@ -0,0 +1,83 @@
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+
+/// Prune trials whose intermediate values exceed fixed thresholds.
+///
+/// Useful for cutting off trials that are clearly diverging or stuck
+/// at bad values early in training.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::pruner::ThresholdPruner;
+///
+/// // Prune if the intermediate value exceeds 100.0 or falls below 0.0
+/// let pruner = ThresholdPruner::new().upper(100.0).lower(0.0);
+/// ```
+pub struct ThresholdPruner {
+    /// Prune if intermediate value is greater than this. `None` = no upper bound.
+    upper: Option<f64>,
+    /// Prune if intermediate value is less than this. `None` = no lower bound.
+    lower: Option<f64>,
+}
+
+impl ThresholdPruner {
+    /// Create a new `ThresholdPruner` with no thresholds set.
+    ///
+    /// By default, no pruning occurs. Use [`upper`](Self::upper) and
+    /// [`lower`](Self::lower) to set bounds.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            upper: None,
+            lower: None,
+        }
+    }
+
+    /// Set the upper threshold. Trials with intermediate values above this
+    /// will be pruned.
+    #[must_use]
+    pub fn upper(mut self, threshold: f64) -> Self {
+        self.upper = Some(threshold);
+        self
+    }
+
+    /// Set the lower threshold. Trials with intermediate values below this
+    /// will be pruned.
+    #[must_use]
+    pub fn lower(mut self, threshold: f64) -> Self {
+        self.lower = Some(threshold);
+        self
+    }
+}
+
+impl Default for ThresholdPruner {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl Pruner for ThresholdPruner {
+    fn should_prune(
+        &self,
+        _trial_id: u64,
+        _step: u64,
+        intermediate_values: &[(u64, f64)],
+        _completed_trials: &[CompletedTrial],
+    ) -> bool {
+        let Some(&(_, latest_value)) = intermediate_values.last() else {
+            return false;
+        };
+        if let Some(upper) = self.upper
+            && latest_value > upper
+        {
+            return true;
+        }
+        if let Some(lower) = self.lower
+            && latest_value < lower
+        {
+            return true;
+        }
+        false
+    }
+}

src/pruner/wilcoxon.rs

@@ -0,0 +1,531 @@
+use core::cmp::Ordering;
+
+use super::Pruner;
+use crate::sampler::CompletedTrial;
+use crate::types::{Direction, TrialState};
+
+/// Prune trials using a Wilcoxon signed-rank test comparing intermediate
+/// values against the best completed trial.
+///
+/// More principled than `MedianPruner` for noisy objectives — it accounts
+/// for the paired nature of step-aligned comparisons and doesn't prune
+/// on random fluctuations.
+///
+/// The test compares intermediate values at matching steps between the
+/// current trial and the best completed trial. If the current trial is
+/// statistically significantly worse (p < threshold), it is pruned.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::pruner::WilcoxonPruner;
+///
+/// let pruner = WilcoxonPruner::new(Direction::Minimize)
+///     .p_value_threshold(0.05)
+///     .n_warmup_steps(5)
+///     .n_min_trials(1);
+/// ```
+pub struct WilcoxonPruner {
+    /// Significance level (default 0.05). Lower = more conservative.
+    p_value_threshold: f64,
+    /// Don't prune in the first N steps (let the trial warm up).
+    n_warmup_steps: u64,
+    /// Require at least N completed trials before pruning.
+    n_min_trials: usize,
+    /// The optimization direction.
+    direction: Direction,
+}
+
+impl WilcoxonPruner {
+    /// Create a new `WilcoxonPruner` for the given optimization direction.
+    ///
+    /// By default, `p_value_threshold` is 0.05, `n_warmup_steps` is 0,
+    /// and `n_min_trials` is 1.
+    #[must_use]
+    pub fn new(direction: Direction) -> Self {
+        Self {
+            p_value_threshold: 0.05,
+            n_warmup_steps: 0,
+            n_min_trials: 1,
+            direction,
+        }
+    }
+
+    /// Set the p-value threshold for significance.
+    ///
+    /// Must be in (0.0, 1.0). Lower values are more conservative (harder to prune).
+    ///
+    /// # Panics
+    ///
+    /// Panics if `p` is not in the open interval (0.0, 1.0).
+    #[must_use]
+    pub fn p_value_threshold(mut self, p: f64) -> Self {
+        assert!(
+            p > 0.0 && p < 1.0,
+            "p_value_threshold must be in (0.0, 1.0)"
+        );
+        self.p_value_threshold = p;
+        self
+    }
+
+    /// Set the number of warmup steps. No pruning occurs before this step.
+    #[must_use]
+    pub fn n_warmup_steps(mut self, n: u64) -> Self {
+        self.n_warmup_steps = n;
+        self
+    }
+
+    /// Set the minimum number of completed trials required before pruning.
+    #[must_use]
+    pub fn n_min_trials(mut self, n: usize) -> Self {
+        self.n_min_trials = n;
+        self
+    }
+}
+
+impl Pruner for WilcoxonPruner {
+    fn should_prune(
+        &self,
+        _trial_id: u64,
+        step: u64,
+        intermediate_values: &[(u64, f64)],
+        completed_trials: &[CompletedTrial],
+    ) -> bool {
+        if step < self.n_warmup_steps {
+            return false;
+        }
+
+        let completed: Vec<&CompletedTrial> = completed_trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .collect();
+
+        if completed.len() < self.n_min_trials {
+            return false;
+        }
+
+        // Find the best completed trial by final objective value.
+        let best = match self.direction {
+            Direction::Minimize => completed
+                .iter()
+                .min_by(|a, b| a.value.partial_cmp(&b.value).unwrap_or(Ordering::Equal)),
+            Direction::Maximize => completed
+                .iter()
+                .max_by(|a, b| a.value.partial_cmp(&b.value).unwrap_or(Ordering::Equal)),
+        };
+
+        let Some(best) = best else {
+            return false;
+        };
+
+        // Pair intermediate values at matching steps.
+        let pairs: Vec<(f64, f64)> = intermediate_values
+            .iter()
+            .filter_map(|&(s, current_v)| {
+                best.intermediate_values
+                    .iter()
+                    .find(|(bs, _)| *bs == s)
+                    .map(|&(_, best_v)| (current_v, best_v))
+            })
+            .collect();
+
+        // Need at least 6 pairs for a meaningful test.
+        if pairs.len() < 6 {
+            return false;
+        }
+
+        // Compute signed differences: current - best.
+        // For minimization: positive diff means current is worse.
+        // For maximization: negative diff means current is worse.
+        let differences: Vec<f64> = pairs
+            .iter()
+            .map(|&(current, best_v)| current - best_v)
+            .collect();
+
+        // Run the Wilcoxon signed-rank test.
+        let p_value = wilcoxon_signed_rank_test(&differences, self.direction);
+
+        p_value < self.p_value_threshold
+    }
+}
+
+/// Perform a one-sided Wilcoxon signed-rank test.
+///
+/// Tests whether the values tend to be worse than zero (positive for
+/// minimization, negative for maximization).
+///
+/// Returns a p-value. Small p-values indicate the current trial is
+/// significantly worse.
+fn wilcoxon_signed_rank_test(differences: &[f64], direction: Direction) -> f64 {
+    // 1. Remove zero differences.
+    let nonzero: Vec<f64> = differences.iter().copied().filter(|d| *d != 0.0).collect();
+    let n = nonzero.len();
+
+    if n < 6 {
+        return 1.0; // Not enough data
+    }
+
+    // 2. Rank by absolute value.
+    let mut abs_ranked: Vec<(usize, f64, f64)> = nonzero
+        .iter()
+        .enumerate()
+        .map(|(i, &d)| (i, d.abs(), d))
+        .collect();
+    abs_ranked.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(Ordering::Equal));
+
+    // 3. Assign ranks with tie correction.
+    let ranks = assign_ranks(&abs_ranked);
+
+    // 4. Compute W+ (sum of ranks for positive differences) and
+    //    W- (sum of ranks for negative differences).
+    let mut w_plus = 0.0;
+    let mut w_minus = 0.0;
+    for (i, &(_, _, orig)) in abs_ranked.iter().enumerate() {
+        if orig > 0.0 {
+            w_plus += ranks[i];
+        } else {
+            w_minus += ranks[i];
+        }
+    }
+
+    // For one-sided test:
+    // - Minimization: we want to detect positive diffs (current worse).
+    //   Large W+ means significantly worse. Test statistic = W-.
+    // - Maximization: we want to detect negative diffs (current worse).
+    //   Large W- means significantly worse. Test statistic = W+.
+    let w = match direction {
+        Direction::Minimize => w_minus,
+        Direction::Maximize => w_plus,
+    };
+
+    // 5. Normal approximation for the p-value.
+    #[allow(clippy::cast_precision_loss)]
+    let n_f = n as f64;
+    let mean = n_f * (n_f + 1.0) / 4.0;
+    let variance = n_f * (n_f + 1.0) * (2.0 * n_f + 1.0) / 24.0;
+
+    // Tie correction for variance.
+    let tie_correction = compute_tie_correction(&ranks);
+    let adjusted_variance = variance - tie_correction;
+
+    if adjusted_variance <= 0.0 {
+        return 1.0;
+    }
+
+    let std_dev = adjusted_variance.sqrt();
+
+    // Continuity correction: shift W by 0.5 towards mean.
+    let continuity = if w < mean { 0.5 } else { -0.5 };
+    let z = (w + continuity - mean) / std_dev;
+
+    // One-sided p-value (lower tail): probability that the test statistic
+    // is this small or smaller under H0.
+    normal_cdf(z)
+}
+
+/// Assign average ranks, handling ties.
+fn assign_ranks(sorted: &[(usize, f64, f64)]) -> Vec<f64> {
+    let n = sorted.len();
+    let mut ranks = vec![0.0; n];
+    let mut i = 0;
+
+    while i < n {
+        let mut j = i;
+        // Find all items tied with sorted[i].
+        while j < n
+            && (sorted[j].1 - sorted[i].1).abs() < f64::EPSILON * sorted[i].1.max(1.0) * 100.0
+        {
+            j += 1;
+        }
+        // Average rank for the tie group. Ranks are 1-based.
+        #[allow(clippy::cast_precision_loss)]
+        let avg_rank = (i + 1 + j) as f64 / 2.0;
+        for rank in ranks.iter_mut().take(j).skip(i) {
+            *rank = avg_rank;
+        }
+        i = j;
+    }
+
+    ranks
+}
+
+/// Compute the tie correction term for the variance.
+/// For each tie group of size t, subtract t^3 - t from the sum,
+/// then divide by 48.
+fn compute_tie_correction(ranks: &[f64]) -> f64 {
+    let mut correction = 0.0;
+    let mut i = 0;
+
+    while i < ranks.len() {
+        let mut j = i;
+        while j < ranks.len() && (ranks[j] - ranks[i]).abs() < f64::EPSILON {
+            j += 1;
+        }
+        #[allow(clippy::cast_precision_loss)]
+        let t = (j - i) as f64;
+        if t > 1.0 {
+            correction += t * t * t - t;
+        }
+        i = j;
+    }
+
+    correction / 48.0
+}
+
+/// Standard normal CDF using an approximation (Abramowitz & Stegun).
+fn normal_cdf(x: f64) -> f64 {
+    // Use the complementary error function relationship:
+    // Φ(x) = 0.5 * erfc(-x / √2)
+    0.5 * erfc(-x / core::f64::consts::SQRT_2)
+}
+
+/// Complementary error function approximation.
+/// Maximum error: 1.5 × 10⁻⁷ (Abramowitz & Stegun formula 7.1.26).
+fn erfc(x: f64) -> f64 {
+    let t = 1.0 / (1.0 + 0.327_591_1 * x.abs());
+    let poly = t
+        * (0.254_829_592
+            + t * (-0.284_496_736
+                + t * (1.421_413_741 + t * (-1.453_152_027 + t * 1.061_405_429))));
+    let result = poly * (-x * x).exp();
+
+    if x >= 0.0 { result } else { 2.0 - result }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+
+    use super::*;
+
+    fn trial_with_values(
+        id: u64,
+        value: f64,
+        intermediate_values: Vec<(u64, f64)>,
+    ) -> CompletedTrial {
+        CompletedTrial::with_intermediate_values(
+            id,
+            HashMap::new(),
+            HashMap::new(),
+            HashMap::new(),
+            value,
+            intermediate_values,
+            HashMap::new(),
+        )
+    }
+
+    #[test]
+    fn no_prune_during_warmup() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize).n_warmup_steps(10);
+        let completed = vec![trial_with_values(
+            0,
+            0.1,
+            (0..20).map(|s| (s, 0.1)).collect(),
+        )];
+        let current: Vec<(u64, f64)> = (0..8).map(|s| (s, 100.0)).collect();
+        assert!(!pruner.should_prune(1, 7, &current, &completed));
+    }
+
+    #[test]
+    fn no_prune_with_insufficient_trials() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize).n_min_trials(5);
+        let completed = vec![trial_with_values(
+            0,
+            0.1,
+            (0..20).map(|s| (s, 0.1)).collect(),
+        )];
+        let current: Vec<(u64, f64)> = (0..10).map(|s| (s, 100.0)).collect();
+        assert!(!pruner.should_prune(1, 9, &current, &completed));
+    }
+
+    #[test]
+    fn no_prune_with_fewer_than_6_pairs() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+        let completed = vec![trial_with_values(
+            0,
+            0.1,
+            (0..5).map(|s| (s, 0.1)).collect(),
+        )];
+        // Only 5 matching steps
+        let current: Vec<(u64, f64)> = (0..5).map(|s| (s, 100.0)).collect();
+        assert!(!pruner.should_prune(1, 4, &current, &completed));
+    }
+
+    #[test]
+    fn prune_when_consistently_worse_minimize() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+
+        // Best trial has low values.
+        let best_values: Vec<(u64, f64)> = (0..20).map(|s| (s, 0.1)).collect();
+        let completed = vec![trial_with_values(0, 0.1, best_values)];
+
+        // Current trial is consistently much worse.
+        let current: Vec<(u64, f64)> = (0..20).map(|s| (s, 10.0)).collect();
+        assert!(pruner.should_prune(1, 19, &current, &completed));
+    }
+
+    #[test]
+    fn prune_when_consistently_worse_maximize() {
+        let pruner = WilcoxonPruner::new(Direction::Maximize);
+
+        // Best trial has high values.
+        let best_values: Vec<(u64, f64)> = (0..20).map(|s| (s, 10.0)).collect();
+        let completed = vec![trial_with_values(0, 10.0, best_values)];
+
+        // Current trial is consistently much worse.
+        let current: Vec<(u64, f64)> = (0..20).map(|s| (s, 0.1)).collect();
+        assert!(pruner.should_prune(1, 19, &current, &completed));
+    }
+
+    #[test]
+    fn no_prune_when_statistically_similar() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+
+        // Best trial and current trial have very similar values with noise.
+        let best_values: Vec<(u64, f64)> = (0..20_u64)
+            .map(|s| {
+                let noise = if s.is_multiple_of(2) { 0.01 } else { -0.01 };
+                (s, 1.0 + noise)
+            })
+            .collect();
+        let completed = vec![trial_with_values(0, 1.0, best_values)];
+
+        // Current trial is similar — alternating above/below.
+        let current: Vec<(u64, f64)> = (0..20_u64)
+            .map(|s| {
+                let noise = if s.is_multiple_of(2) { -0.01 } else { 0.01 };
+                (s, 1.0 + noise)
+            })
+            .collect();
+        assert!(!pruner.should_prune(1, 19, &current, &completed));
+    }
+
+    #[test]
+    fn selects_best_trial_minimize() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+
+        // Two completed trials: trial 0 is better (lower).
+        let completed = vec![
+            trial_with_values(0, 0.1, (0..20).map(|s| (s, 0.1)).collect()),
+            trial_with_values(1, 5.0, (0..20).map(|s| (s, 5.0)).collect()),
+        ];
+
+        // Current trial is worse than the best but similar to the second.
+        let current: Vec<(u64, f64)> = (0..20).map(|s| (s, 5.0)).collect();
+        assert!(pruner.should_prune(2, 19, &current, &completed));
+    }
+
+    #[test]
+    fn selects_best_trial_maximize() {
+        let pruner = WilcoxonPruner::new(Direction::Maximize);
+
+        // Two completed trials: trial 1 is better (higher).
+        let completed = vec![
+            trial_with_values(0, 0.1, (0..20).map(|s| (s, 0.1)).collect()),
+            trial_with_values(1, 10.0, (0..20).map(|s| (s, 10.0)).collect()),
+        ];
+
+        // Current trial is worse than the best.
+        let current: Vec<(u64, f64)> = (0..20).map(|s| (s, 0.1)).collect();
+        assert!(pruner.should_prune(2, 19, &current, &completed));
+    }
+
+    #[test]
+    fn ignores_pruned_trials() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+
+        // Only a pruned trial — no complete trials.
+        let mut trial = trial_with_values(0, 0.1, (0..20).map(|s| (s, 0.1)).collect());
+        trial.state = TrialState::Pruned;
+        let completed = vec![trial];
+
+        let current: Vec<(u64, f64)> = (0..20).map(|s| (s, 100.0)).collect();
+        assert!(!pruner.should_prune(1, 19, &current, &completed));
+    }
+
+    #[test]
+    fn lower_p_value_is_more_conservative() {
+        let strict = WilcoxonPruner::new(Direction::Minimize).p_value_threshold(0.001);
+        let lenient = WilcoxonPruner::new(Direction::Minimize).p_value_threshold(0.1);
+
+        let completed = vec![trial_with_values(
+            0,
+            0.1,
+            (0..20).map(|s| (s, 0.1)).collect(),
+        )];
+
+        // Moderately worse — should pass lenient but maybe not strict.
+        let current: Vec<(u64, f64)> = (0..20)
+            .map(|s| if s < 15 { (s, 0.2) } else { (s, 0.15) })
+            .collect();
+
+        let lenient_prunes = lenient.should_prune(1, 19, &current, &completed);
+        let strict_prunes = strict.should_prune(1, 19, &current, &completed);
+
+        // A stricter threshold should never prune when a lenient one doesn't.
+        if !lenient_prunes {
+            assert!(!strict_prunes);
+        }
+    }
+
+    #[test]
+    #[should_panic(expected = "p_value_threshold must be in (0.0, 1.0)")]
+    fn panics_on_zero_p_value() {
+        let _ = WilcoxonPruner::new(Direction::Minimize).p_value_threshold(0.0);
+    }
+
+    #[test]
+    #[should_panic(expected = "p_value_threshold must be in (0.0, 1.0)")]
+    fn panics_on_one_p_value() {
+        let _ = WilcoxonPruner::new(Direction::Minimize).p_value_threshold(1.0);
+    }
+
+    #[test]
+    fn correct_signed_rank_statistic() {
+        // Known example: differences [1, 2, 3, 4, 5, 6] (all positive).
+        // Ranks: 1, 2, 3, 4, 5, 6. W+ = 21, W- = 0.
+        // For minimization (testing if positive = worse), W- = 0.
+        // This should give a very small p-value.
+        let diffs = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
+        let p = wilcoxon_signed_rank_test(&diffs, Direction::Minimize);
+        assert!(
+            p < 0.05,
+            "p-value {p} should be < 0.05 for all-positive diffs"
+        );
+    }
+
+    #[test]
+    fn symmetric_differences_not_significant() {
+        // Balanced differences: half positive, half negative.
+        let diffs = vec![1.0, -1.0, 2.0, -2.0, 3.0, -3.0, 4.0, -4.0];
+        let p = wilcoxon_signed_rank_test(&diffs, Direction::Minimize);
+        assert!(p > 0.05, "p-value {p} should be > 0.05 for symmetric diffs");
+    }
+
+    #[test]
+    fn normal_cdf_known_values() {
+        assert!((normal_cdf(0.0) - 0.5).abs() < 1e-6);
+        assert!(normal_cdf(-10.0) < 1e-6);
+        assert!((normal_cdf(10.0) - 1.0).abs() < 1e-6);
+        assert!((normal_cdf(-1.96) - 0.025).abs() < 0.001);
+    }
+
+    #[test]
+    fn no_intermediate_values() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+        let completed = vec![trial_with_values(
+            0,
+            0.1,
+            (0..20).map(|s| (s, 0.1)).collect(),
+        )];
+        assert!(!pruner.should_prune(1, 0, &[], &completed));
+    }
+
+    #[test]
+    fn no_completed_trials() {
+        let pruner = WilcoxonPruner::new(Direction::Minimize);
+        let current: Vec<(u64, f64)> = (0..20).map(|s| (s, 1.0)).collect();
+        assert!(!pruner.should_prune(1, 19, &current, &[]));
+    }
+}

src/sampler/mod.rs

@@ -9,6 +9,8 @@ use std::collections::HashMap;
 use crate::distribution::Distribution;
 use crate::param::ParamValue;
 use crate::parameter::{ParamId, Parameter};
+use crate::trial::AttrValue;
+use crate::types::TrialState;
 
 /// A completed trial with its parameters, distributions, and objective value.
 ///
@@ -27,6 +29,12 @@ pub struct CompletedTrial<V = f64> {
     pub param_labels: HashMap<ParamId, String>,
     /// The objective value returned by the objective function.
     pub value: V,
+    /// Intermediate objective values reported during the trial.
+    pub intermediate_values: Vec<(u64, f64)>,
+    /// The state of the trial (Complete, Pruned, or Failed).
+    pub state: TrialState,
+    /// User-defined attributes stored during the trial.
+    pub user_attrs: HashMap<String, AttrValue>,
 }
 
 impl<V> CompletedTrial<V> {
@@ -44,6 +52,31 @@ impl<V> CompletedTrial<V> {
             distributions,
             param_labels,
             value,
+            intermediate_values: Vec::new(),
+            state: TrialState::Complete,
+            user_attrs: HashMap::new(),
+        }
+    }
+
+    /// Creates a new completed trial with intermediate values and user attributes.
+    pub fn with_intermediate_values(
+        id: u64,
+        params: HashMap<ParamId, ParamValue>,
+        distributions: HashMap<ParamId, Distribution>,
+        param_labels: HashMap<ParamId, String>,
+        value: V,
+        intermediate_values: Vec<(u64, f64)>,
+        user_attrs: HashMap<String, AttrValue>,
+    ) -> Self {
+        Self {
+            id,
+            params,
+            distributions,
+            param_labels,
+            value,
+            intermediate_values,
+            state: TrialState::Complete,
+            user_attrs,
         }
     }
 
@@ -87,6 +120,18 @@ impl<V> CompletedTrial<V> {
                 .expect("parameter type mismatch: stored value incompatible with parameter")
         })
     }
+
+    /// Gets a user attribute by key.
+    #[must_use]
+    pub fn user_attr(&self, key: &str) -> Option<&AttrValue> {
+        self.user_attrs.get(key)
+    }
+
+    /// Returns all user attributes.
+    #[must_use]
+    pub fn user_attrs(&self) -> &HashMap<String, AttrValue> {
+        &self.user_attrs
+    }
 }
 
 /// A pending (running) trial with its parameters and distributions, but no objective value yet.

src/study.rs

@@ -5,14 +5,20 @@ use core::any::Any;
 use core::future::Future;
 use core::ops::ControlFlow;
 use core::sync::atomic::{AtomicU64, Ordering};
+use core::time::Duration;
+use std::collections::{HashMap, VecDeque};
 use std::sync::Arc;
+use std::time::Instant;
 
-use parking_lot::RwLock;
+use parking_lot::{Mutex, RwLock};
 
+use crate::param::ParamValue;
+use crate::parameter::ParamId;
+use crate::pruner::{NopPruner, Pruner};
 use crate::sampler::random::RandomSampler;
 use crate::sampler::{CompletedTrial, Sampler};
 use crate::trial::Trial;
-use crate::types::Direction;
+use crate::types::{Direction, TrialState};
 
 /// A study manages the optimization process, tracking trials and their results.
 ///
@@ -40,6 +46,8 @@ where
     direction: Direction,
     /// The sampler used to generate parameter values.
     sampler: Arc<dyn Sampler>,
+    /// The pruner used to decide whether to stop trials early.
+    pruner: Arc<dyn Pruner>,
     /// Completed trials (wrapped in Arc for sharing with Trial).
     completed_trials: Arc<RwLock<Vec<CompletedTrial<V>>>>,
     /// Counter for generating unique trial IDs.
@@ -48,6 +56,8 @@ where
     /// Set automatically for `Study<f64>` so that `create_trial()` and all
     /// optimization methods use the sampler without requiring `_with_sampler` suffixes.
     trial_factory: Option<Arc<dyn Fn(u64) -> Trial + Send + Sync>>,
+    /// Queue of parameter configurations to evaluate next.
+    enqueued_params: Arc<Mutex<VecDeque<HashMap<ParamId, ParamValue>>>>,
 }
 
 impl<V> Study<V>
@@ -78,6 +88,56 @@ where
         Self::with_sampler(direction, RandomSampler::new())
     }
 
+    /// Creates a study that minimizes the objective value.
+    ///
+    /// This is a shorthand for `Study::with_sampler(Direction::Minimize, sampler)`.
+    ///
+    /// # Arguments
+    ///
+    /// * `sampler` - The sampler to use for parameter sampling.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Study;
+    /// use optimizer::sampler::tpe::TpeSampler;
+    ///
+    /// let study: Study<f64> = Study::minimize(TpeSampler::new());
+    /// assert_eq!(study.direction(), optimizer::Direction::Minimize);
+    /// ```
+    #[must_use]
+    pub fn minimize(sampler: impl Sampler + 'static) -> Self
+    where
+        V: 'static,
+    {
+        Self::with_sampler(Direction::Minimize, sampler)
+    }
+
+    /// Creates a study that maximizes the objective value.
+    ///
+    /// This is a shorthand for `Study::with_sampler(Direction::Maximize, sampler)`.
+    ///
+    /// # Arguments
+    ///
+    /// * `sampler` - The sampler to use for parameter sampling.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Study;
+    /// use optimizer::sampler::tpe::TpeSampler;
+    ///
+    /// let study: Study<f64> = Study::maximize(TpeSampler::new());
+    /// assert_eq!(study.direction(), optimizer::Direction::Maximize);
+    /// ```
+    #[must_use]
+    pub fn maximize(sampler: impl Sampler + 'static) -> Self
+    where
+        V: 'static,
+    {
+        Self::with_sampler(Direction::Maximize, sampler)
+    }
+
     /// Creates a new study with a custom sampler.
     ///
     /// # Arguments
@@ -102,16 +162,20 @@ where
         let sampler: Arc<dyn Sampler> = Arc::new(sampler);
         let completed_trials = Arc::new(RwLock::new(Vec::new()));
 
+        let pruner: Arc<dyn Pruner> = Arc::new(NopPruner);
+
         // For Study<f64>, set up a trial factory that provides sampler integration.
         // This uses Any downcasting to check at runtime whether V = f64.
-        let trial_factory = Self::make_trial_factory(&sampler, &completed_trials);
+        let trial_factory = Self::make_trial_factory(&sampler, &completed_trials, &pruner);
 
         Self {
             direction,
             sampler,
+            pruner,
             completed_trials,
             next_trial_id: AtomicU64::new(0),
             trial_factory,
+            enqueued_params: Arc::new(Mutex::new(VecDeque::new())),
         }
     }
 
@@ -119,6 +183,7 @@ where
     fn make_trial_factory(
         sampler: &Arc<dyn Sampler>,
         completed_trials: &Arc<RwLock<Vec<CompletedTrial<V>>>>,
+        pruner: &Arc<dyn Pruner>,
     ) -> Option<Arc<dyn Fn(u64) -> Trial + Send + Sync>>
     where
         V: 'static,
@@ -131,8 +196,14 @@ where
         f64_trials.map(|trials| {
             let sampler = Arc::clone(sampler);
             let trials = Arc::clone(trials);
+            let pruner = Arc::clone(pruner);
             let factory: Arc<dyn Fn(u64) -> Trial + Send + Sync> = Arc::new(move |id| {
-                Trial::with_sampler(id, Arc::clone(&sampler), Arc::clone(&trials))
+                Trial::with_sampler(
+                    id,
+                    Arc::clone(&sampler),
+                    Arc::clone(&trials),
+                    Arc::clone(&pruner),
+                )
             });
             factory
         })
@@ -158,12 +229,120 @@ where
     /// let mut study: Study<f64> = Study::new(Direction::Minimize);
     /// study.set_sampler(TpeSampler::new());
     /// ```
+    /// Creates a new study with a custom sampler and pruner.
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    /// * `sampler` - The sampler to use for parameter sampling.
+    /// * `pruner` - The pruner to use for trial pruning.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::pruner::NopPruner;
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler_and_pruner(Direction::Minimize, sampler, NopPruner);
+    /// ```
+    pub fn with_sampler_and_pruner(
+        direction: Direction,
+        sampler: impl Sampler + 'static,
+        pruner: impl Pruner + 'static,
+    ) -> Self
+    where
+        V: 'static,
+    {
+        let sampler: Arc<dyn Sampler> = Arc::new(sampler);
+        let pruner: Arc<dyn Pruner> = Arc::new(pruner);
+        let completed_trials = Arc::new(RwLock::new(Vec::new()));
+        let trial_factory = Self::make_trial_factory(&sampler, &completed_trials, &pruner);
+
+        Self {
+            direction,
+            sampler,
+            pruner,
+            completed_trials,
+            next_trial_id: AtomicU64::new(0),
+            trial_factory,
+            enqueued_params: Arc::new(Mutex::new(VecDeque::new())),
+        }
+    }
+
     pub fn set_sampler(&mut self, sampler: impl Sampler + 'static)
     where
         V: 'static,
     {
         self.sampler = Arc::new(sampler);
-        self.trial_factory = Self::make_trial_factory(&self.sampler, &self.completed_trials);
+        self.trial_factory =
+            Self::make_trial_factory(&self.sampler, &self.completed_trials, &self.pruner);
+    }
+
+    /// Sets a new pruner for the study.
+    ///
+    /// # Arguments
+    ///
+    /// * `pruner` - The pruner to use for trial pruning.
+    pub fn set_pruner(&mut self, pruner: impl Pruner + 'static)
+    where
+        V: 'static,
+    {
+        self.pruner = Arc::new(pruner);
+        self.trial_factory =
+            Self::make_trial_factory(&self.sampler, &self.completed_trials, &self.pruner);
+    }
+
+    /// Returns a reference to the study's pruner.
+    pub fn pruner(&self) -> &dyn Pruner {
+        &*self.pruner
+    }
+
+    /// Enqueues a specific parameter configuration to be evaluated next.
+    ///
+    /// The next call to [`ask()`](Self::ask) or the next trial in [`optimize()`](Self::optimize)
+    /// will use these exact parameters instead of sampling from the sampler.
+    ///
+    /// Multiple configurations can be enqueued; they are evaluated in FIFO order.
+    /// If an enqueued configuration is missing a parameter that the objective calls
+    /// `suggest()` on, that parameter falls back to normal sampling.
+    ///
+    /// # Arguments
+    ///
+    /// * `params` - A map from parameter IDs to the values to use.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use std::collections::HashMap;
+    ///
+    /// use optimizer::parameter::{FloatParam, IntParam, Parameter};
+    /// use optimizer::{Direction, ParamValue, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0);
+    /// let y = IntParam::new(1, 100);
+    ///
+    /// // Evaluate these specific configurations first
+    /// study.enqueue(HashMap::from([
+    ///     (x.id(), ParamValue::Float(0.001)),
+    ///     (y.id(), ParamValue::Int(3)),
+    /// ]));
+    ///
+    /// // Next trial will use x=0.001, y=3
+    /// let mut trial = study.ask();
+    /// assert_eq!(x.suggest(&mut trial).unwrap(), 0.001);
+    /// assert_eq!(y.suggest(&mut trial).unwrap(), 3);
+    /// ```
+    pub fn enqueue(&self, params: HashMap<ParamId, ParamValue>) {
+        self.enqueued_params.lock().push_back(params);
+    }
+
+    /// Returns the number of enqueued parameter configurations.
+    #[must_use]
+    pub fn n_enqueued(&self) -> usize {
+        self.enqueued_params.lock().len()
     }
 
     /// Generates the next unique trial ID.
@@ -195,11 +374,18 @@ where
     /// ```
     pub fn create_trial(&self) -> Trial {
         let id = self.next_trial_id();
-        if let Some(factory) = &self.trial_factory {
+        let mut trial = if let Some(factory) = &self.trial_factory {
             factory(id)
         } else {
             Trial::new(id)
+        };
+
+        // If there are enqueued params, inject them into this trial
+        if let Some(fixed_params) = self.enqueued_params.lock().pop_front() {
+            trial.set_fixed_params(fixed_params);
         }
+
+        trial
     }
 
     /// Records a completed trial with its objective value.
@@ -230,13 +416,16 @@ where
     /// ```
     pub fn complete_trial(&self, mut trial: Trial, value: V) {
         trial.set_complete();
-        let completed = CompletedTrial::new(
+        let mut completed = CompletedTrial::with_intermediate_values(
             trial.id(),
             trial.params().clone(),
             trial.distributions().clone(),
             trial.param_labels().clone(),
             value,
+            trial.intermediate_values().to_vec(),
+            trial.user_attrs().clone(),
         );
+        completed.state = TrialState::Complete;
         self.completed_trials.write().push(completed);
     }
 
@@ -270,6 +459,85 @@ where
         // They could be stored in a separate list for debugging if needed
     }
 
+    /// Request a new trial with suggested parameters.
+    ///
+    /// This is the first half of the ask-and-tell interface. After calling
+    /// `ask()`, use parameter types to suggest values on the returned trial,
+    /// evaluate your objective externally, then pass the trial back to
+    /// [`tell()`](Self::tell) with the result.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0);
+    ///
+    /// let mut trial = study.ask();
+    /// let x_val = x.suggest(&mut trial).unwrap();
+    /// let value = x_val * x_val;
+    /// study.tell(trial, Ok::<_, &str>(value));
+    /// ```
+    pub fn ask(&self) -> Trial {
+        self.create_trial()
+    }
+
+    /// Report the result of a trial obtained from [`ask()`](Self::ask).
+    ///
+    /// Pass `Ok(value)` for a successful evaluation or `Err(reason)` for a
+    /// failure. Failed trials are not stored in the study's history.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    ///
+    /// let trial = study.ask();
+    /// study.tell(trial, Ok::<_, &str>(42.0));
+    /// assert_eq!(study.n_trials(), 1);
+    ///
+    /// let trial = study.ask();
+    /// study.tell(trial, Err::<f64, _>("evaluation failed"));
+    /// assert_eq!(study.n_trials(), 1); // failed trials not counted
+    /// ```
+    pub fn tell(&self, trial: Trial, value: core::result::Result<V, impl ToString>) {
+        match value {
+            Ok(v) => self.complete_trial(trial, v),
+            Err(e) => self.fail_trial(trial, e),
+        }
+    }
+
+    /// Records a pruned trial, preserving its intermediate values.
+    ///
+    /// Pruned trials are stored alongside completed trials so that samplers
+    /// can optionally learn from partial evaluations. The trial's state is
+    /// set to `Pruned`.
+    ///
+    /// # Arguments
+    ///
+    /// * `trial` - The trial that was pruned.
+    pub fn prune_trial(&self, mut trial: Trial)
+    where
+        V: Default,
+    {
+        trial.set_pruned();
+        let mut completed = CompletedTrial::with_intermediate_values(
+            trial.id(),
+            trial.params().clone(),
+            trial.distributions().clone(),
+            trial.param_labels().clone(),
+            V::default(),
+            trial.intermediate_values().to_vec(),
+            trial.user_attrs().clone(),
+        );
+        completed.state = TrialState::Pruned;
+        self.completed_trials.write().push(completed);
+    }
+
     /// Returns an iterator over all completed trials.
     ///
     /// The iterator yields references to `CompletedTrial` values, which contain
@@ -324,6 +592,15 @@ where
         self.completed_trials.read().len()
     }
 
+    /// Returns the number of pruned trials.
+    pub fn n_pruned_trials(&self) -> usize {
+        self.completed_trials
+            .read()
+            .iter()
+            .filter(|t| t.state == TrialState::Pruned)
+            .count()
+    }
+
     /// Returns the trial with the best objective value.
     ///
     /// The "best" trial depends on the optimization direction:
@@ -364,12 +641,9 @@ where
     {
         let trials = self.completed_trials.read();
 
-        if trials.is_empty() {
-            return Err(crate::Error::NoCompletedTrials);
-        }
-
         let best = trials
             .iter()
+            .filter(|t| t.state == TrialState::Complete)
             .max_by(|a, b| {
                 // For Minimize, we want the smallest value to be "max" in ordering
                 // For Maximize, we want the largest value to be "max" in ordering
@@ -431,6 +705,37 @@ where
         self.best_trial().map(|trial| trial.value)
     }
 
+    /// Returns the top `n` trials sorted by objective value.
+    ///
+    /// For `Direction::Minimize`, returns trials with the lowest values.
+    /// For `Direction::Maximize`, returns trials with the highest values.
+    /// Only includes completed trials (not failed or pruned).
+    ///
+    /// If fewer than `n` completed trials exist, returns all of them.
+    pub fn top_trials(&self, n: usize) -> Vec<CompletedTrial<V>>
+    where
+        V: Clone,
+    {
+        let trials = self.completed_trials.read();
+        let mut completed: Vec<_> = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .cloned()
+            .collect();
+        completed.sort_by(|a, b| match self.direction {
+            Direction::Minimize => a
+                .value
+                .partial_cmp(&b.value)
+                .unwrap_or(core::cmp::Ordering::Equal),
+            Direction::Maximize => b
+                .value
+                .partial_cmp(&a.value)
+                .unwrap_or(core::cmp::Ordering::Equal),
+        });
+        completed.truncate(n);
+        completed
+    }
+
     /// Runs optimization with the given objective function.
     ///
     /// This method runs `n_trials` evaluations sequentially. For each trial:
@@ -480,7 +785,8 @@ where
     pub fn optimize<F, E>(&self, n_trials: usize, mut objective: F) -> crate::Result<()>
     where
         F: FnMut(&mut Trial) -> core::result::Result<V, E>,
-        E: ToString,
+        E: ToString + 'static,
+        V: Default,
     {
         for _ in 0..n_trials {
             let mut trial = self.create_trial();
@@ -490,13 +796,22 @@ where
                     self.complete_trial(trial, value);
                 }
                 Err(e) => {
-                    self.fail_trial(trial, e.to_string());
+                    if is_trial_pruned(&e) {
+                        self.prune_trial(trial);
+                    } else {
+                        self.fail_trial(trial, e.to_string());
+                    }
                 }
             }
         }
 
-        // Return error if no trials succeeded
-        if self.n_trials() == 0 {
+        // Return error if no trials completed successfully
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
             return Err(crate::Error::NoCompletedTrials);
         }
 
@@ -581,8 +896,13 @@ where
             }
         }
 
-        // Return error if no trials succeeded
-        if self.n_trials() == 0 {
+        // Return error if no trials completed successfully
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
             return Err(crate::Error::NoCompletedTrials);
         }
 
@@ -696,8 +1016,13 @@ where
             }
         }
 
-        // Return error if no trials succeeded
-        if self.n_trials() == 0 {
+        // Return error if no trials completed successfully
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
             return Err(crate::Error::NoCompletedTrials);
         }
 
@@ -768,10 +1093,10 @@ where
         mut callback: C,
     ) -> crate::Result<()>
     where
-        V: Clone,
+        V: Clone + Default,
         F: FnMut(&mut Trial) -> core::result::Result<V, E>,
         C: FnMut(&Study<V>, &CompletedTrial<V>) -> ControlFlow<()>,
-        E: ToString,
+        E: ToString + 'static,
     {
         for _ in 0..n_trials {
             let mut trial = self.create_trial();
@@ -799,13 +1124,340 @@ where
                     }
                 }
                 Err(e) => {
-                    self.fail_trial(trial, e.to_string());
+                    if is_trial_pruned(&e) {
+                        self.prune_trial(trial);
+                    } else {
+                        self.fail_trial(trial, e.to_string());
+                    }
                 }
             }
         }
 
-        // Return error if no trials succeeded
-        if self.n_trials() == 0 {
+        // Return error if no trials completed successfully
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+    /// Runs optimization until the given duration has elapsed.
+    ///
+    /// Trials that are already running when the timeout is reached will
+    /// complete — we never interrupt mid-trial. The actual elapsed time
+    /// may therefore slightly exceed the specified duration.
+    ///
+    /// # Arguments
+    ///
+    /// * `duration` - The maximum wall-clock time to spend on optimization.
+    /// * `objective` - A closure that takes a mutable reference to a `Trial` and
+    ///   returns the objective value or an error.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully
+    /// before the timeout.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use std::time::Duration;
+    ///
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    ///
+    /// let x_param = FloatParam::new(-10.0, 10.0);
+    ///
+    /// study
+    ///     .optimize_until(Duration::from_millis(100), |trial| {
+    ///         let x = x_param.suggest(trial)?;
+    ///         Ok::<_, optimizer::Error>(x * x)
+    ///     })
+    ///     .unwrap();
+    ///
+    /// assert!(study.n_trials() > 0);
+    /// ```
+    pub fn optimize_until<F, E>(&self, duration: Duration, mut objective: F) -> crate::Result<()>
+    where
+        F: FnMut(&mut Trial) -> core::result::Result<V, E>,
+        E: ToString + 'static,
+        V: Default,
+    {
+        let deadline = Instant::now() + duration;
+        while Instant::now() < deadline {
+            let mut trial = self.create_trial();
+
+            match objective(&mut trial) {
+                Ok(value) => {
+                    self.complete_trial(trial, value);
+                }
+                Err(e) => {
+                    if is_trial_pruned(&e) {
+                        self.prune_trial(trial);
+                    } else {
+                        self.fail_trial(trial, e.to_string());
+                    }
+                }
+            }
+        }
+
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+
+    /// Runs optimization until the given duration has elapsed, with a callback.
+    ///
+    /// Like [`optimize_until`](Self::optimize_until), but calls a callback after
+    /// each completed trial. The callback can stop optimization early by returning
+    /// `ControlFlow::Break(())`.
+    ///
+    /// # Arguments
+    ///
+    /// * `duration` - The maximum wall-clock time to spend on optimization.
+    /// * `objective` - A closure that takes a mutable reference to a `Trial` and
+    ///   returns the objective value or an error.
+    /// * `callback` - A closure called after each successful trial. Returns
+    ///   `ControlFlow::Continue(())` to proceed or `ControlFlow::Break(())` to stop.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully.
+    /// Returns `Error::Internal` if a completed trial is not found after adding.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use std::ops::ControlFlow;
+    /// use std::time::Duration;
+    ///
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    ///
+    /// let x_param = FloatParam::new(-10.0, 10.0);
+    ///
+    /// study
+    ///     .optimize_until_with_callback(
+    ///         Duration::from_secs(1),
+    ///         |trial| {
+    ///             let x = x_param.suggest(trial)?;
+    ///             Ok::<_, optimizer::Error>(x * x)
+    ///         },
+    ///         |_study, completed_trial| {
+    ///             if completed_trial.value < 1.0 {
+    ///                 ControlFlow::Break(())
+    ///             } else {
+    ///                 ControlFlow::Continue(())
+    ///             }
+    ///         },
+    ///     )
+    ///     .unwrap();
+    ///
+    /// assert!(study.n_trials() > 0);
+    /// ```
+    pub fn optimize_until_with_callback<F, C, E>(
+        &self,
+        duration: Duration,
+        mut objective: F,
+        mut callback: C,
+    ) -> crate::Result<()>
+    where
+        V: Clone + Default,
+        F: FnMut(&mut Trial) -> core::result::Result<V, E>,
+        C: FnMut(&Study<V>, &CompletedTrial<V>) -> ControlFlow<()>,
+        E: ToString + 'static,
+    {
+        let deadline = Instant::now() + duration;
+        while Instant::now() < deadline {
+            let mut trial = self.create_trial();
+
+            match objective(&mut trial) {
+                Ok(value) => {
+                    self.complete_trial(trial, value);
+
+                    let trials = self.completed_trials.read();
+                    let Some(completed) = trials.last() else {
+                        return Err(crate::Error::Internal(
+                            "completed trial not found after adding",
+                        ));
+                    };
+
+                    let completed_clone = completed.clone();
+                    drop(trials);
+
+                    if let ControlFlow::Break(()) = callback(self, &completed_clone) {
+                        break;
+                    }
+                }
+                Err(e) => {
+                    if is_trial_pruned(&e) {
+                        self.prune_trial(trial);
+                    } else {
+                        self.fail_trial(trial, e.to_string());
+                    }
+                }
+            }
+        }
+
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+
+    /// Runs optimization asynchronously until the given duration has elapsed.
+    ///
+    /// The async variant of [`optimize_until`](Self::optimize_until). Trials are
+    /// run sequentially, but the objective function can be async.
+    ///
+    /// # Arguments
+    ///
+    /// * `duration` - The maximum wall-clock time to spend on optimization.
+    /// * `objective` - A function that takes a `Trial` and returns a `Future`
+    ///   that resolves to a tuple of `(Trial, Result<V, E>)`.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully.
+    #[cfg(feature = "async")]
+    pub async fn optimize_until_async<F, Fut, E>(
+        &self,
+        duration: Duration,
+        objective: F,
+    ) -> crate::Result<()>
+    where
+        F: Fn(Trial) -> Fut,
+        Fut: Future<Output = core::result::Result<(Trial, V), E>>,
+        E: ToString,
+    {
+        let deadline = Instant::now() + duration;
+        while Instant::now() < deadline {
+            let trial = self.create_trial();
+
+            match objective(trial).await {
+                Ok((trial, value)) => {
+                    self.complete_trial(trial, value);
+                }
+                Err(e) => {
+                    let _ = e.to_string();
+                }
+            }
+        }
+
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+
+    /// Runs optimization with bounded parallelism until the given duration has elapsed.
+    ///
+    /// The parallel variant of [`optimize_until`](Self::optimize_until). Runs up to
+    /// `concurrency` trials simultaneously using async tasks. New trials are spawned
+    /// as long as the deadline has not been reached; trials already running when the
+    /// deadline passes will complete.
+    ///
+    /// # Arguments
+    ///
+    /// * `duration` - The maximum wall-clock time to spend spawning new trials.
+    /// * `concurrency` - The maximum number of trials to run simultaneously.
+    /// * `objective` - A function that takes a `Trial` and returns a `Future`
+    ///   that resolves to a tuple of `(Trial, V)` or an error.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully.
+    /// Returns `Error::TaskError` if the semaphore is closed or a spawned task panics.
+    #[cfg(feature = "async")]
+    pub async fn optimize_until_parallel<F, Fut, E>(
+        &self,
+        duration: Duration,
+        concurrency: usize,
+        objective: F,
+    ) -> crate::Result<()>
+    where
+        F: Fn(Trial) -> Fut + Send + Sync + 'static,
+        Fut: Future<Output = core::result::Result<(Trial, V), E>> + Send,
+        E: ToString + Send + 'static,
+        V: Send + 'static,
+    {
+        use tokio::sync::Semaphore;
+
+        let deadline = Instant::now() + duration;
+        let semaphore = Arc::new(Semaphore::new(concurrency));
+        let objective = Arc::new(objective);
+
+        let mut handles = Vec::new();
+
+        while Instant::now() < deadline {
+            let permit = semaphore
+                .clone()
+                .acquire_owned()
+                .await
+                .map_err(|e| crate::Error::TaskError(e.to_string()))?;
+            let trial = self.create_trial();
+            let objective = Arc::clone(&objective);
+
+            let handle = tokio::spawn(async move {
+                let result = objective(trial).await;
+                drop(permit);
+                result
+            });
+
+            handles.push(handle);
+        }
+
+        for handle in handles {
+            match handle
+                .await
+                .map_err(|e| crate::Error::TaskError(e.to_string()))?
+            {
+                Ok((trial, value)) => {
+                    self.complete_trial(trial, value);
+                }
+                Err(e) => {
+                    let _ = e.to_string();
+                }
+            }
+        }
+
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
             return Err(crate::Error::NoCompletedTrials);
         }
 
@@ -843,7 +1495,7 @@ impl Study<f64> {
     pub fn optimize_with_sampler<F, E>(&self, n_trials: usize, objective: F) -> crate::Result<()>
     where
         F: FnMut(&mut Trial) -> core::result::Result<f64, E>,
-        E: ToString,
+        E: ToString + 'static,
     {
         self.optimize(n_trials, objective)
     }
@@ -865,7 +1517,7 @@ impl Study<f64> {
     where
         F: FnMut(&mut Trial) -> core::result::Result<f64, E>,
         C: FnMut(&Study<f64>, &CompletedTrial<f64>) -> ControlFlow<()>,
-        E: ToString,
+        E: ToString + 'static,
     {
         self.optimize_with_callback(n_trials, objective, callback)
     }
@@ -916,3 +1568,16 @@ impl Study<f64> {
             .await
     }
 }
+
+/// Returns `true` if the error represents a pruned trial.
+///
+/// Checks via `Any` downcasting whether `e` is `Error::TrialPruned` or
+/// the standalone `TrialPruned` struct.
+fn is_trial_pruned<E: 'static>(e: &E) -> bool {
+    let any: &dyn Any = e;
+    if let Some(err) = any.downcast_ref::<crate::Error>() {
+        matches!(err, crate::Error::TrialPruned)
+    } else {
+        any.downcast_ref::<crate::error::TrialPruned>().is_some()
+    }
+}

src/trial.rs

@@ -9,9 +9,53 @@ use crate::distribution::Distribution;
 use crate::error::{Error, Result};
 use crate::param::ParamValue;
 use crate::parameter::{ParamId, Parameter};
+use crate::pruner::Pruner;
 use crate::sampler::{CompletedTrial, Sampler};
 use crate::types::TrialState;
 
+/// A user attribute value that can be stored on a trial.
+#[derive(Clone, Debug, PartialEq)]
+pub enum AttrValue {
+    /// A floating-point attribute.
+    Float(f64),
+    /// An integer attribute.
+    Int(i64),
+    /// A string attribute.
+    String(String),
+    /// A boolean attribute.
+    Bool(bool),
+}
+
+impl From<f64> for AttrValue {
+    fn from(v: f64) -> Self {
+        Self::Float(v)
+    }
+}
+
+impl From<i64> for AttrValue {
+    fn from(v: i64) -> Self {
+        Self::Int(v)
+    }
+}
+
+impl From<String> for AttrValue {
+    fn from(v: String) -> Self {
+        Self::String(v)
+    }
+}
+
+impl From<&str> for AttrValue {
+    fn from(v: &str) -> Self {
+        Self::String(v.to_owned())
+    }
+}
+
+impl From<bool> for AttrValue {
+    fn from(v: bool) -> Self {
+        Self::Bool(v)
+    }
+}
+
 /// A trial represents a single evaluation of the objective function.
 ///
 /// Each trial has a unique ID and stores the sampled parameters along with
@@ -36,6 +80,14 @@ pub struct Trial {
     sampler: Option<Arc<dyn Sampler>>,
     /// Access to the history of completed trials (shared with Study).
     history: Option<Arc<RwLock<Vec<CompletedTrial<f64>>>>>,
+    /// Intermediate objective values reported at each step.
+    intermediate_values: Vec<(u64, f64)>,
+    /// The pruner used to decide whether to stop this trial early.
+    pruner: Option<Arc<dyn Pruner>>,
+    /// User-defined attributes for logging, debugging, and analysis.
+    user_attrs: HashMap<String, AttrValue>,
+    /// Pre-filled parameter values from enqueue (used instead of sampling).
+    fixed_params: HashMap<ParamId, ParamValue>,
 }
 
 impl core::fmt::Debug for Trial {
@@ -48,6 +100,10 @@ impl core::fmt::Debug for Trial {
             .field("param_labels", &self.param_labels)
             .field("has_sampler", &self.sampler.is_some())
             .field("has_history", &self.history.is_some())
+            .field("intermediate_values", &self.intermediate_values)
+            .field("has_pruner", &self.pruner.is_some())
+            .field("user_attrs", &self.user_attrs)
+            .field("fixed_params", &self.fixed_params)
             .finish()
     }
 }
@@ -83,6 +139,10 @@ impl Trial {
             param_labels: HashMap::new(),
             sampler: None,
             history: None,
+            intermediate_values: Vec::new(),
+            pruner: None,
+            user_attrs: HashMap::new(),
+            fixed_params: HashMap::new(),
         }
     }
 
@@ -100,6 +160,7 @@ impl Trial {
         id: u64,
         sampler: Arc<dyn Sampler>,
         history: Arc<RwLock<Vec<CompletedTrial<f64>>>>,
+        pruner: Arc<dyn Pruner>,
     ) -> Self {
         Self {
             id,
@@ -109,9 +170,21 @@ impl Trial {
             param_labels: HashMap::new(),
             sampler: Some(sampler),
             history: Some(history),
+            intermediate_values: Vec::new(),
+            pruner: Some(pruner),
+            user_attrs: HashMap::new(),
+            fixed_params: HashMap::new(),
         }
     }
 
+    /// Sets pre-filled parameters on this trial.
+    ///
+    /// When `suggest_param` is called for a parameter that has a fixed value,
+    /// the fixed value is used instead of sampling.
+    pub(crate) fn set_fixed_params(&mut self, params: HashMap<ParamId, ParamValue>) {
+        self.fixed_params = params;
+    }
+
     /// Samples a value from the given distribution using the sampler.
     ///
     /// If the trial has a sampler, it delegates to the sampler's sample method
@@ -159,6 +232,61 @@ impl Trial {
         &self.param_labels
     }
 
+    /// Reports an intermediate objective value at a given step.
+    ///
+    /// Steps should be monotonically increasing (e.g., epoch number).
+    /// Duplicate steps overwrite the previous value.
+    pub fn report(&mut self, step: u64, value: f64) {
+        if let Some(entry) = self
+            .intermediate_values
+            .iter_mut()
+            .find(|(s, _)| *s == step)
+        {
+            entry.1 = value;
+        } else {
+            self.intermediate_values.push((step, value));
+        }
+    }
+
+    /// Ask whether this trial should be pruned at the current step.
+    ///
+    /// Returns `true` if the pruner recommends stopping this trial.
+    /// The caller should return `Err(TrialPruned)` from the objective.
+    #[must_use]
+    pub fn should_prune(&self) -> bool {
+        let (Some(pruner), Some(history)) = (&self.pruner, &self.history) else {
+            return false;
+        };
+        let Some(&(step, _)) = self.intermediate_values.last() else {
+            return false;
+        };
+        let history_guard = history.read();
+        pruner.should_prune(self.id, step, &self.intermediate_values, &history_guard)
+    }
+
+    /// Returns all intermediate values reported so far.
+    #[must_use]
+    pub fn intermediate_values(&self) -> &[(u64, f64)] {
+        &self.intermediate_values
+    }
+
+    /// Sets a user attribute on this trial.
+    pub fn set_user_attr(&mut self, key: impl Into<String>, value: impl Into<AttrValue>) {
+        self.user_attrs.insert(key.into(), value.into());
+    }
+
+    /// Gets a user attribute by key.
+    #[must_use]
+    pub fn user_attr(&self, key: &str) -> Option<&AttrValue> {
+        self.user_attrs.get(key)
+    }
+
+    /// Returns all user attributes.
+    #[must_use]
+    pub fn user_attrs(&self) -> &HashMap<String, AttrValue> {
+        &self.user_attrs
+    }
+
     /// Sets the trial state to Complete.
     pub(crate) fn set_complete(&mut self) {
         self.state = TrialState::Complete;
@@ -169,6 +297,11 @@ impl Trial {
         self.state = TrialState::Failed;
     }
 
+    /// Sets the trial state to Pruned.
+    pub(crate) fn set_pruned(&mut self) {
+        self.state = TrialState::Pruned;
+    }
+
     /// Suggests a parameter value using a [`Parameter`] definition.
     ///
     /// This is the primary entry point for sampling parameters. It handles
@@ -223,8 +356,14 @@ impl Trial {
             });
         }
 
-        // Sample using the sampler
-        let value = self.sample_value(&distribution);
+        // Check for a pre-filled (enqueued) value for this parameter
+        let value = if let Some(fixed_value) = self.fixed_params.remove(&param_id) {
+            fixed_value
+        } else {
+            // Sample using the sampler
+            self.sample_value(&distribution)
+        };
+
         let result = param.cast_param_value(&value)?;
 
         // Store distribution, value, and label

src/types.rs

@@ -18,4 +18,6 @@ pub enum TrialState {
     Complete,
     /// The trial failed with an error.
     Failed,
+    /// The trial was pruned (stopped early).
+    Pruned,
 }

tests/integration.rs

@@ -1363,3 +1363,532 @@ fn test_completed_trial_get() {
     assert!((-10.0..=10.0).contains(&x_val));
     assert!((1..=10).contains(&n_val));
 }
+
+// =============================================================================
+// Tests for timeout-based optimization
+// =============================================================================
+
+#[test]
+fn test_optimize_until_runs_for_approximately_specified_duration() {
+    use std::time::{Duration, Instant};
+
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x_param = FloatParam::new(-10.0, 10.0);
+
+    let duration = Duration::from_millis(200);
+    let start = Instant::now();
+
+    study
+        .optimize_until(duration, |trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x * x)
+        })
+        .unwrap();
+
+    let elapsed = start.elapsed();
+    assert!(
+        elapsed >= duration,
+        "should run for at least the specified duration, elapsed: {elapsed:?}"
+    );
+    // Allow generous upper bound — the last trial may overshoot
+    assert!(
+        elapsed < duration + Duration::from_millis(200),
+        "should not overshoot excessively, elapsed: {elapsed:?}"
+    );
+}
+
+#[test]
+fn test_optimize_until_completes_at_least_one_trial() {
+    use std::time::Duration;
+
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x_param = FloatParam::new(-10.0, 10.0);
+
+    study
+        .optimize_until(Duration::from_millis(100), |trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x * x)
+        })
+        .unwrap();
+
+    assert!(
+        study.n_trials() >= 1,
+        "should complete at least one trial, got {}",
+        study.n_trials()
+    );
+}
+
+#[test]
+fn test_optimize_until_works_with_minimize() {
+    use std::time::Duration;
+
+    let sampler = RandomSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    let x_param = FloatParam::new(-10.0, 10.0);
+
+    study
+        .optimize_until(Duration::from_millis(100), |trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x * x)
+        })
+        .unwrap();
+
+    let best = study.best_value().unwrap();
+    assert!(best >= 0.0, "x^2 should be non-negative");
+}
+
+#[test]
+fn test_optimize_until_works_with_maximize() {
+    use std::time::Duration;
+
+    let sampler = RandomSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Maximize, sampler);
+    let x_param = FloatParam::new(0.0, 10.0);
+
+    study
+        .optimize_until(Duration::from_millis(100), |trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x)
+        })
+        .unwrap();
+
+    let best = study.best_value().unwrap();
+    assert!(best >= 0.0);
+}
+
+#[test]
+fn test_optimize_until_with_callback_early_stopping() {
+    use std::ops::ControlFlow;
+    use std::time::Duration;
+
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x_param = FloatParam::new(0.0, 10.0);
+
+    study
+        .optimize_until_with_callback(
+            Duration::from_secs(10), // long timeout — callback should stop early
+            |trial| {
+                let x = x_param.suggest(trial)?;
+                Ok::<_, Error>(x)
+            },
+            |study, _trial| {
+                if study.n_trials() >= 5 {
+                    ControlFlow::Break(())
+                } else {
+                    ControlFlow::Continue(())
+                }
+            },
+        )
+        .unwrap();
+
+    assert_eq!(
+        study.n_trials(),
+        5,
+        "callback should have stopped after 5 trials"
+    );
+}
+
+#[test]
+fn test_optimize_until_all_trials_fail() {
+    use std::time::Duration;
+
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    let result = study.optimize_until(Duration::from_millis(50), |_trial| {
+        Err::<f64, &str>("always fails")
+    });
+
+    assert!(
+        matches!(result, Err(Error::NoCompletedTrials)),
+        "should return NoCompletedTrials when all trials fail"
+    );
+}
+
+#[test]
+fn test_optimize_until_with_non_f64_value_type() {
+    use std::time::Duration;
+
+    let study: Study<i32> = Study::new(Direction::Minimize);
+    let x_param = IntParam::new(-10, 10);
+
+    study
+        .optimize_until(Duration::from_millis(100), |trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x.abs() as i32)
+        })
+        .unwrap();
+
+    assert!(study.n_trials() >= 1);
+    let best = study.best_trial().unwrap();
+    assert!(best.value >= 0);
+}
+
+// =============================================================================
+// Tests for top_trials
+// =============================================================================
+
+#[test]
+fn test_top_trials_minimize() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    // Manually complete trials with known values
+    for &val in &[5.0, 1.0, 3.0, 2.0, 4.0] {
+        let trial = study.create_trial();
+        study.complete_trial(trial, val);
+    }
+
+    let top3 = study.top_trials(3);
+    assert_eq!(top3.len(), 3);
+    assert_eq!(top3[0].value, 1.0);
+    assert_eq!(top3[1].value, 2.0);
+    assert_eq!(top3[2].value, 3.0);
+}
+
+#[test]
+fn test_top_trials_maximize() {
+    let study: Study<f64> = Study::new(Direction::Maximize);
+
+    for &val in &[5.0, 1.0, 3.0, 2.0, 4.0] {
+        let trial = study.create_trial();
+        study.complete_trial(trial, val);
+    }
+
+    let top3 = study.top_trials(3);
+    assert_eq!(top3.len(), 3);
+    assert_eq!(top3[0].value, 5.0);
+    assert_eq!(top3[1].value, 4.0);
+    assert_eq!(top3[2].value, 3.0);
+}
+
+#[test]
+fn test_top_trials_n_greater_than_total() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    for &val in &[3.0, 1.0] {
+        let trial = study.create_trial();
+        study.complete_trial(trial, val);
+    }
+
+    let top = study.top_trials(10);
+    assert_eq!(top.len(), 2);
+    assert_eq!(top[0].value, 1.0);
+    assert_eq!(top[1].value, 3.0);
+}
+
+#[test]
+fn test_top_trials_empty() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    let top = study.top_trials(5);
+    assert!(top.is_empty());
+}
+
+#[test]
+fn test_top_trials_excludes_pruned() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    // Complete some trials
+    for &val in &[5.0, 1.0, 3.0] {
+        let trial = study.create_trial();
+        study.complete_trial(trial, val);
+    }
+
+    // Prune a trial (it gets a default value of 0.0 but should be excluded)
+    let trial = study.create_trial();
+    study.prune_trial(trial);
+
+    let top = study.top_trials(5);
+    assert_eq!(top.len(), 3, "pruned trial should be excluded");
+    assert_eq!(top[0].value, 1.0);
+}
+
+// =============================================================================
+// Test: ask-and-tell interface
+// =============================================================================
+
+#[test]
+fn test_ask_and_tell_basic() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x_param = FloatParam::new(0.0, 10.0);
+
+    for _ in 0..10 {
+        let mut trial = study.ask();
+        let x = x_param.suggest(&mut trial).unwrap();
+        let value = x * x;
+        study.tell(trial, Ok::<_, &str>(value));
+    }
+
+    assert_eq!(study.n_trials(), 10);
+    assert!(study.best_value().unwrap() >= 0.0);
+}
+
+#[test]
+fn test_ask_and_tell_with_failures() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x_param = FloatParam::new(-5.0, 5.0);
+
+    // Alternate success and failure
+    for i in 0..10 {
+        let mut trial = study.ask();
+        let x = x_param.suggest(&mut trial).unwrap();
+        if i % 2 == 0 {
+            study.tell(trial, Ok::<_, &str>(x * x));
+        } else {
+            study.tell(trial, Err::<f64, _>("simulated failure"));
+        }
+    }
+
+    // Only successful trials are counted
+    assert_eq!(study.n_trials(), 5);
+}
+
+#[test]
+fn test_ask_and_tell_with_tpe_sampler() {
+    let sampler = TpeSampler::builder()
+        .seed(42)
+        .n_startup_trials(5)
+        .build()
+        .unwrap();
+    let study: Study<f64> = Study::minimize(sampler);
+    let x_param = FloatParam::new(-10.0, 10.0);
+
+    for _ in 0..30 {
+        let mut trial = study.ask();
+        let x = x_param.suggest(&mut trial).unwrap();
+        study.tell(trial, Ok::<_, &str>((x - 3.0).powi(2)));
+    }
+
+    assert_eq!(study.n_trials(), 30);
+    assert!(
+        study.best_value().unwrap() < 5.0,
+        "TPE ask-and-tell should find a reasonable value"
+    );
+}
+
+#[test]
+fn test_ask_and_tell_batch() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x_param = FloatParam::new(0.0, 10.0);
+
+    // Ask a batch of trials
+    let batch: Vec<_> = (0..5)
+        .map(|_| {
+            let mut t = study.ask();
+            let x = x_param.suggest(&mut t).unwrap();
+            (t, x)
+        })
+        .collect();
+
+    // Tell results for the batch
+    for (trial, x) in batch {
+        study.tell(trial, Ok::<_, &str>(x * x));
+    }
+
+    assert_eq!(study.n_trials(), 5);
+}
+
+#[test]
+fn test_ask_and_tell_with_custom_value_type() {
+    // Ask-and-tell works with non-f64 value types too
+    let study: Study<i32> = Study::new(Direction::Maximize);
+
+    for i in 0..5 {
+        let trial = study.ask();
+        study.tell(trial, Ok::<_, &str>(i * 10));
+    }
+
+    assert_eq!(study.n_trials(), 5);
+    assert_eq!(study.best_value().unwrap(), 40);
+}
+
+// =============================================================================
+// Tests: enqueue trials
+// =============================================================================
+
+use std::collections::HashMap;
+
+use optimizer::ParamValue;
+
+#[test]
+fn test_enqueue_params_evaluated_first() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+    let y = IntParam::new(1, 100);
+
+    // Enqueue a specific configuration
+    study.enqueue(HashMap::from([
+        (x.id(), ParamValue::Float(5.0)),
+        (y.id(), ParamValue::Int(42)),
+    ]));
+
+    // The first trial should use the enqueued params
+    let mut trial = study.ask();
+    let x_val = x.suggest(&mut trial).unwrap();
+    let y_val = y.suggest(&mut trial).unwrap();
+
+    assert_eq!(x_val, 5.0);
+    assert_eq!(y_val, 42);
+}
+
+#[test]
+fn test_enqueue_fifo_order() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+
+    // Enqueue two configs
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(1.0))]));
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(2.0))]));
+
+    // First trial gets first enqueued value
+    let mut trial1 = study.ask();
+    assert_eq!(x.suggest(&mut trial1).unwrap(), 1.0);
+
+    // Second trial gets second enqueued value
+    let mut trial2 = study.ask();
+    assert_eq!(x.suggest(&mut trial2).unwrap(), 2.0);
+}
+
+#[test]
+fn test_enqueue_then_normal_sampling_resumes() {
+    let sampler = RandomSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    let x = FloatParam::new(0.0, 10.0);
+
+    // Enqueue one config
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(5.0))]));
+
+    // First trial uses enqueued value
+    let mut trial1 = study.ask();
+    assert_eq!(x.suggest(&mut trial1).unwrap(), 5.0);
+    study.tell(trial1, Ok::<_, &str>(25.0));
+
+    // Second trial uses normal sampling (not 5.0)
+    let mut trial2 = study.ask();
+    let x_val = x.suggest(&mut trial2).unwrap();
+    // The sampled value should be in [0, 10] but extremely unlikely to be exactly 5.0
+    assert!((0.0..=10.0).contains(&x_val));
+}
+
+#[test]
+fn test_enqueue_with_optimize() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+
+    // Enqueue two specific configs
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(1.0))]));
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(2.0))]));
+
+    let mut values = Vec::new();
+
+    study
+        .optimize(5, |trial| {
+            let x_val = x.suggest(trial)?;
+            values.push(x_val);
+            Ok::<_, Error>(x_val * x_val)
+        })
+        .unwrap();
+
+    // First two trials should use enqueued values
+    assert_eq!(values[0], 1.0);
+    assert_eq!(values[1], 2.0);
+    // All 5 trials should have completed
+    assert_eq!(study.n_trials(), 5);
+}
+
+#[test]
+fn test_enqueue_partial_params_fall_back_to_sampling() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+    let y = IntParam::new(1, 100);
+
+    // Enqueue only x, not y
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(3.0))]));
+
+    let mut trial = study.ask();
+    let x_val = x.suggest(&mut trial).unwrap();
+    let y_val = y.suggest(&mut trial).unwrap();
+
+    // x should be the enqueued value
+    assert_eq!(x_val, 3.0);
+    // y should be sampled (within range)
+    assert!((1..=100).contains(&y_val));
+}
+
+#[test]
+fn test_enqueue_trials_appear_in_completed_trials() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(7.0))]));
+
+    study
+        .optimize(1, |trial| {
+            let x_val = x.suggest(trial)?;
+            Ok::<_, Error>(x_val)
+        })
+        .unwrap();
+
+    let trials = study.trials();
+    assert_eq!(trials.len(), 1);
+    assert_eq!(trials[0].value, 7.0);
+    assert_eq!(
+        *trials[0].params.get(&x.id()).unwrap(),
+        ParamValue::Float(7.0)
+    );
+}
+
+#[test]
+fn test_enqueue_with_ask_and_tell() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(4.0))]));
+
+    let mut trial = study.ask();
+    let x_val = x.suggest(&mut trial).unwrap();
+    assert_eq!(x_val, 4.0);
+
+    study.tell(trial, Ok::<_, &str>(x_val * x_val));
+    assert_eq!(study.n_trials(), 1);
+    assert_eq!(study.best_value().unwrap(), 16.0);
+}
+
+#[test]
+fn test_n_enqueued() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+
+    assert_eq!(study.n_enqueued(), 0);
+
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(1.0))]));
+    assert_eq!(study.n_enqueued(), 1);
+
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(2.0))]));
+    assert_eq!(study.n_enqueued(), 2);
+
+    // Creating a trial dequeues one
+    let _ = study.ask();
+    assert_eq!(study.n_enqueued(), 1);
+
+    let _ = study.ask();
+    assert_eq!(study.n_enqueued(), 0);
+}
+
+#[test]
+fn test_enqueue_counted_in_n_trials() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 10.0);
+
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(1.0))]));
+    study.enqueue(HashMap::from([(x.id(), ParamValue::Float(2.0))]));
+
+    study
+        .optimize(5, |trial| {
+            let x_val = x.suggest(trial)?;
+            Ok::<_, Error>(x_val)
+        })
+        .unwrap();
+
+    // All 5 trials count, including the 2 enqueued ones
+    assert_eq!(study.n_trials(), 5);
+}

tests/median_pruner_tests.rs

@@ -0,0 +1,228 @@
+use std::collections::HashMap;
+
+use optimizer::Direction;
+use optimizer::pruner::{MedianPruner, Pruner};
+use optimizer::sampler::CompletedTrial;
+
+/// Helper to build a completed trial with given intermediate values.
+fn trial_with_values(id: u64, intermediate_values: Vec<(u64, f64)>) -> CompletedTrial {
+    CompletedTrial::with_intermediate_values(
+        id,
+        HashMap::new(),
+        HashMap::new(),
+        HashMap::new(),
+        0.0,
+        intermediate_values,
+        HashMap::new(),
+    )
+}
+
+// --- Minimize direction ---
+
+#[test]
+fn prune_when_worse_than_median_minimize() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    // 3 completed trials with values at step 2: [1.0, 2.0, 3.0] => median = 2.0
+    let completed = vec![
+        trial_with_values(0, vec![(0, 0.5), (1, 0.8), (2, 1.0)]),
+        trial_with_values(1, vec![(0, 0.6), (1, 1.5), (2, 2.0)]),
+        trial_with_values(2, vec![(0, 0.7), (1, 2.0), (2, 3.0)]),
+    ];
+    // Current trial value at step 2 is 2.5 > median 2.0 => prune
+    let current = vec![(0, 0.5), (1, 1.0), (2, 2.5)];
+    assert!(pruner.should_prune(3, 2, &current, &completed));
+}
+
+#[test]
+fn no_prune_when_better_than_median_minimize() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    let completed = vec![
+        trial_with_values(0, vec![(0, 0.5), (1, 0.8), (2, 1.0)]),
+        trial_with_values(1, vec![(0, 0.6), (1, 1.5), (2, 2.0)]),
+        trial_with_values(2, vec![(0, 0.7), (1, 2.0), (2, 3.0)]),
+    ];
+    // Current trial value at step 2 is 1.5 < median 2.0 => don't prune
+    let current = vec![(0, 0.5), (1, 1.0), (2, 1.5)];
+    assert!(!pruner.should_prune(3, 2, &current, &completed));
+}
+
+// --- Maximize direction ---
+
+#[test]
+fn prune_when_worse_than_median_maximize() {
+    let pruner = MedianPruner::new(Direction::Maximize);
+    // Values at step 1: [5.0, 7.0, 9.0] => median = 7.0
+    let completed = vec![
+        trial_with_values(0, vec![(0, 3.0), (1, 5.0)]),
+        trial_with_values(1, vec![(0, 4.0), (1, 7.0)]),
+        trial_with_values(2, vec![(0, 5.0), (1, 9.0)]),
+    ];
+    // Current value 6.0 < median 7.0 => prune (worse for maximize)
+    let current = vec![(0, 4.0), (1, 6.0)];
+    assert!(pruner.should_prune(3, 1, &current, &completed));
+}
+
+#[test]
+fn no_prune_when_better_than_median_maximize() {
+    let pruner = MedianPruner::new(Direction::Maximize);
+    let completed = vec![
+        trial_with_values(0, vec![(0, 3.0), (1, 5.0)]),
+        trial_with_values(1, vec![(0, 4.0), (1, 7.0)]),
+        trial_with_values(2, vec![(0, 5.0), (1, 9.0)]),
+    ];
+    // Current value 8.0 > median 7.0 => don't prune
+    let current = vec![(0, 4.0), (1, 8.0)];
+    assert!(!pruner.should_prune(3, 1, &current, &completed));
+}
+
+// --- Warmup steps ---
+
+#[test]
+fn no_prune_during_warmup() {
+    let pruner = MedianPruner::new(Direction::Minimize).n_warmup_steps(5);
+    let completed = vec![trial_with_values(0, vec![(0, 1.0), (1, 1.0), (2, 1.0)])];
+    // Step 2 < warmup 5 => never prune, even if value is terrible
+    let current = vec![(0, 100.0), (1, 100.0), (2, 100.0)];
+    assert!(!pruner.should_prune(1, 2, &current, &completed));
+}
+
+#[test]
+fn prune_after_warmup() {
+    let pruner = MedianPruner::new(Direction::Minimize).n_warmup_steps(2);
+    let completed = vec![trial_with_values(0, vec![(0, 1.0), (1, 1.0), (2, 1.0)])];
+    // Step 2 >= warmup 2 => pruning allowed; current 100.0 > median 1.0
+    let current = vec![(0, 100.0), (1, 100.0), (2, 100.0)];
+    assert!(pruner.should_prune(1, 2, &current, &completed));
+}
+
+// --- n_min_trials ---
+
+#[test]
+fn no_prune_when_fewer_than_n_min_trials() {
+    let pruner = MedianPruner::new(Direction::Minimize).n_min_trials(3);
+    // Only 2 completed trials — below threshold of 3
+    let completed = vec![
+        trial_with_values(0, vec![(0, 1.0)]),
+        trial_with_values(1, vec![(0, 2.0)]),
+    ];
+    let current = vec![(0, 100.0)];
+    assert!(!pruner.should_prune(2, 0, &current, &completed));
+}
+
+#[test]
+fn prune_when_at_least_n_min_trials() {
+    let pruner = MedianPruner::new(Direction::Minimize).n_min_trials(3);
+    // 3 completed trials with step 0: [1.0, 2.0, 3.0] => median 2.0
+    let completed = vec![
+        trial_with_values(0, vec![(0, 1.0)]),
+        trial_with_values(1, vec![(0, 2.0)]),
+        trial_with_values(2, vec![(0, 3.0)]),
+    ];
+    // 5.0 > median 2.0 => prune
+    let current = vec![(0, 5.0)];
+    assert!(pruner.should_prune(3, 0, &current, &completed));
+}
+
+// --- No completed trials with values at step ---
+
+#[test]
+fn no_prune_when_no_completed_trials_at_step() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    // Completed trials only have values at step 0, not step 5
+    let completed = vec![
+        trial_with_values(0, vec![(0, 1.0)]),
+        trial_with_values(1, vec![(0, 2.0)]),
+    ];
+    let current = vec![(0, 0.5), (5, 100.0)];
+    assert!(!pruner.should_prune(2, 5, &current, &completed));
+}
+
+// --- Median calculation edge cases ---
+
+#[test]
+fn correct_median_with_even_number_of_trials() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    // 4 trials at step 0: [1.0, 2.0, 3.0, 4.0] => median = 2.5
+    let completed = vec![
+        trial_with_values(0, vec![(0, 1.0)]),
+        trial_with_values(1, vec![(0, 2.0)]),
+        trial_with_values(2, vec![(0, 3.0)]),
+        trial_with_values(3, vec![(0, 4.0)]),
+    ];
+    // 2.6 > 2.5 => prune
+    let current = vec![(0, 2.6)];
+    assert!(pruner.should_prune(4, 0, &current, &completed));
+    // 2.4 < 2.5 => don't prune
+    let current = vec![(0, 2.4)];
+    assert!(!pruner.should_prune(4, 0, &current, &completed));
+}
+
+#[test]
+fn correct_median_with_odd_number_of_trials() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    // 5 trials at step 0: [1.0, 2.0, 3.0, 4.0, 5.0] => median = 3.0
+    let completed = vec![
+        trial_with_values(0, vec![(0, 1.0)]),
+        trial_with_values(1, vec![(0, 2.0)]),
+        trial_with_values(2, vec![(0, 3.0)]),
+        trial_with_values(3, vec![(0, 4.0)]),
+        trial_with_values(4, vec![(0, 5.0)]),
+    ];
+    // 3.5 > 3.0 => prune
+    let current = vec![(0, 3.5)];
+    assert!(pruner.should_prune(5, 0, &current, &completed));
+    // 2.5 < 3.0 => don't prune
+    let current = vec![(0, 2.5)];
+    assert!(!pruner.should_prune(5, 0, &current, &completed));
+}
+
+// --- Non-contiguous step numbers ---
+
+#[test]
+fn works_with_non_contiguous_steps() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    // Steps are 0, 10, 100 — non-contiguous
+    let completed = vec![
+        trial_with_values(0, vec![(0, 1.0), (10, 2.0), (100, 3.0)]),
+        trial_with_values(1, vec![(0, 1.5), (10, 2.5), (100, 4.0)]),
+        trial_with_values(2, vec![(0, 2.0), (10, 3.0), (100, 5.0)]),
+    ];
+    // At step 100: [3.0, 4.0, 5.0] => median = 4.0
+    let current = vec![(0, 1.0), (10, 2.0), (100, 4.5)];
+    assert!(pruner.should_prune(3, 100, &current, &completed));
+
+    let current = vec![(0, 1.0), (10, 2.0), (100, 3.5)];
+    assert!(!pruner.should_prune(3, 100, &current, &completed));
+}
+
+// --- No intermediate values for current trial ---
+
+#[test]
+fn no_prune_when_no_intermediate_values() {
+    let pruner = MedianPruner::new(Direction::Minimize);
+    let completed = vec![trial_with_values(0, vec![(0, 1.0)])];
+    assert!(!pruner.should_prune(1, 0, &[], &completed));
+}
+
+// --- Pruned trials are excluded from median calculation ---
+
+#[test]
+fn pruned_trials_excluded_from_median() {
+    use optimizer::TrialState;
+
+    let pruner = MedianPruner::new(Direction::Minimize);
+
+    let mut pruned = trial_with_values(0, vec![(0, 0.1)]);
+    pruned.state = TrialState::Pruned;
+
+    // Only the completed trial (value 5.0) counts. Pruned trial (0.1) is excluded.
+    let completed = vec![pruned, trial_with_values(1, vec![(0, 5.0)])];
+
+    // 3.0 < 5.0 => don't prune (only 1 completed trial with median 5.0)
+    let current = vec![(0, 3.0)];
+    assert!(!pruner.should_prune(2, 0, &current, &completed));
+
+    // 6.0 > 5.0 => prune
+    let current = vec![(0, 6.0)];
+    assert!(pruner.should_prune(2, 0, &current, &completed));
+}

tests/threshold_pruner_tests.rs

@@ -0,0 +1,64 @@
+use optimizer::pruner::{Pruner, ThresholdPruner};
+
+#[test]
+fn prune_when_value_exceeds_upper_threshold() {
+    let pruner = ThresholdPruner::new().upper(10.0);
+    let values = vec![(0, 5.0), (1, 8.0), (2, 11.0)];
+    assert!(pruner.should_prune(0, 2, &values, &[]));
+}
+
+#[test]
+fn prune_when_value_falls_below_lower_threshold() {
+    let pruner = ThresholdPruner::new().lower(0.0);
+    let values = vec![(0, 5.0), (1, 2.0), (2, -1.0)];
+    assert!(pruner.should_prune(0, 2, &values, &[]));
+}
+
+#[test]
+fn no_prune_when_value_within_bounds() {
+    let pruner = ThresholdPruner::new().upper(10.0).lower(0.0);
+    let values = vec![(0, 3.0), (1, 5.0), (2, 7.0)];
+    assert!(!pruner.should_prune(0, 2, &values, &[]));
+}
+
+#[test]
+fn no_prune_when_no_intermediate_values() {
+    let pruner = ThresholdPruner::new().upper(10.0).lower(0.0);
+    assert!(!pruner.should_prune(0, 0, &[], &[]));
+}
+
+#[test]
+fn works_with_only_upper_set() {
+    let pruner = ThresholdPruner::new().upper(5.0);
+    let below = vec![(0, 3.0)];
+    let above = vec![(0, 6.0)];
+    assert!(!pruner.should_prune(0, 0, &below, &[]));
+    assert!(pruner.should_prune(0, 0, &above, &[]));
+}
+
+#[test]
+fn works_with_only_lower_set() {
+    let pruner = ThresholdPruner::new().lower(2.0);
+    let above = vec![(0, 5.0)];
+    let below = vec![(0, 1.0)];
+    assert!(!pruner.should_prune(0, 0, &above, &[]));
+    assert!(pruner.should_prune(0, 0, &below, &[]));
+}
+
+#[test]
+fn works_with_both_thresholds_set() {
+    let pruner = ThresholdPruner::new().upper(10.0).lower(0.0);
+
+    // Within bounds
+    assert!(!pruner.should_prune(0, 0, &[(0, 5.0)], &[]));
+
+    // Exceeds upper
+    assert!(pruner.should_prune(0, 0, &[(0, 15.0)], &[]));
+
+    // Below lower
+    assert!(pruner.should_prune(0, 0, &[(0, -3.0)], &[]));
+
+    // At exact boundary (not pruned — strictly greater/less)
+    assert!(!pruner.should_prune(0, 0, &[(0, 10.0)], &[]));
+    assert!(!pruner.should_prune(0, 0, &[(0, 0.0)], &[]));
+}

tests/user_attr_tests.rs

@@ -0,0 +1,148 @@
+use optimizer::parameter::{FloatParam, Parameter};
+use optimizer::{AttrValue, Direction, Study};
+
+#[test]
+fn set_and_get_float_attr() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("score", 42.5);
+            assert_eq!(trial.user_attr("score"), Some(&AttrValue::Float(42.5)));
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+}
+
+#[test]
+fn set_and_get_int_attr() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("epoch", 42_i64);
+            assert_eq!(trial.user_attr("epoch"), Some(&AttrValue::Int(42)));
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+}
+
+#[test]
+fn set_and_get_string_attr() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("model", "resnet50");
+            assert_eq!(
+                trial.user_attr("model"),
+                Some(&AttrValue::String("resnet50".to_owned()))
+            );
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+}
+
+#[test]
+fn set_and_get_bool_attr() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("converged", true);
+            assert_eq!(trial.user_attr("converged"), Some(&AttrValue::Bool(true)));
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+}
+
+#[test]
+fn attrs_propagate_to_completed_trial() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("time_secs", 1.5);
+            trial.set_user_attr("tag", "baseline");
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert_eq!(best.user_attr("time_secs"), Some(&AttrValue::Float(1.5)));
+    assert_eq!(
+        best.user_attr("tag"),
+        Some(&AttrValue::String("baseline".to_owned()))
+    );
+}
+
+#[test]
+fn overwrite_attr_replaces_value() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("key", "old");
+            trial.set_user_attr("key", "new");
+            assert_eq!(
+                trial.user_attr("key"),
+                Some(&AttrValue::String("new".to_owned()))
+            );
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert_eq!(
+        best.user_attr("key"),
+        Some(&AttrValue::String("new".to_owned()))
+    );
+}
+
+#[test]
+fn missing_attr_returns_none() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            assert_eq!(trial.user_attr("nonexistent"), None);
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert_eq!(best.user_attr("nonexistent"), None);
+}
+
+#[test]
+fn user_attrs_map_returns_all() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(1, |trial| {
+            let _ = x.suggest(trial)?;
+            trial.set_user_attr("a", 1.0);
+            trial.set_user_attr("b", true);
+            assert_eq!(trial.user_attrs().len(), 2);
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert_eq!(best.user_attrs().len(), 2);
+}