Create multiple PODs where resource limits for a single POD are exceeded (#444)

* Create multiple PODs where resource limits for a single POD are exceeded

* HashSet -> BTreeSet determinism fix

* Fixed incorrect assignment of input PODs and added test

* Ensure only a single output POD

* Return error when reveal() called with unknown statement

* Use unreachable! for presumed-impossible cases

* Use assert_eq! rather than debug_assert_eq

* Use FIFO for topological sort

* Simplify bounds calculation

* Some more simplifications/comments

* Enforce dep_idx < idx invariant

* Incrementally solve rather than estimating slack

* Fix tests to correctly test dependencies between private and public statements

* More tidying

* Note possible optimisation of MainPodBuilder cloning of input PODs

* Fix tracking of total input POD count

* Refactor tests

* Formatting

* Small optimisation: use Vec in place of BTreeSet

* Account for automatically-inserted Contains statements

* Formatting

* Fix possible issue with copied statements

* Simplify result type given only a single result MainPod

* Remove unnecessary POD count estimate functionality

* Simplify dependency ordering and tracking

* Remove notion of multiple output PODs from solver

* Minor simplifications

* Use add_constraint instead of with

* Remove unnecessary check following assertion

* Fix handling of anchored keys given that Contains statements are not auto-inserted if they already exist

* Fix confusing dependency graph test

* Remove prove_order

* Fix deduplication and possible double-counting of public but not copied statements

* Reorder so that the output POD is the final POD

* Add more detailed tests

* Remove redundant tests

* Simplify POD counting

* More docs

* Flag more branches as unreachable

* Formatting

* Fix for changed custom batch parsing

Cargo.toml

@@ -41,6 +41,7 @@ serde_bytes = "0.11"
 serde_arrays = "0.2.0"
 sha2 = { version = "0.10.9" }
 rand_chacha = "0.3.1"
+good_lp = { version = "1.8", default-features = false, features = ["microlp"] }
 
 # Uncomment for debugging with https://github.com/ed255/plonky2/ at branch `feat/debug`.  The repo directory needs to be checked out next to the pod2 repo directory.
 # [patch."https://github.com/0xPARC/plonky2"]

src/frontend/mod.rs

@@ -21,11 +21,15 @@ use crate::middleware::{
 
 mod custom;
 mod error;
+mod multi_pod;
 mod operation;
 mod pod_request;
 mod serialization;
 pub use custom::*;
 pub use error::*;
+pub use multi_pod::{
+    MultiPodBuilder, MultiPodResult, MultiPodSolution, Options as MultiPodOptions,
+};
 pub use operation::*;
 pub use pod_request::*;
 

src/frontend/multi_pod/cost.rs

@@ -0,0 +1,224 @@
+//! Resource cost analysis for statements and operations.
+//!
+//! This module provides cost analysis for multi-POD packing. Each operation
+//! consumes various resources that have per-POD limits.
+
+use std::collections::BTreeSet;
+
+use crate::{
+    frontend::{Operation, OperationArg},
+    middleware::{
+        CustomPredicateBatch, Hash, NativeOperation, OperationType, RawValue, Statement, ValueRef,
+    },
+};
+
+/// Unique identifier for a custom predicate batch.
+///
+/// Uses the batch's cryptographic hash as identifier. Two batches with the same
+/// hash are considered identical for resource counting purposes.
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct CustomBatchId(pub Hash);
+
+impl From<&CustomPredicateBatch> for CustomBatchId {
+    fn from(batch: &CustomPredicateBatch) -> Self {
+        Self(batch.id())
+    }
+}
+
+/// Unique identifier for an anchored key (dict, key) pair.
+///
+/// When a Contains statement is used as an argument to operations like gt(), eq(), etc.,
+/// the value is accessed via an "anchored key" - a reference to a specific key in a
+/// specific dictionary. Each unique anchored key used in a POD requires a Contains
+/// statement to be present in that POD (auto-inserted by MainPodBuilder if needed).
+///
+/// We use the raw values of the dict and key for comparison, as they uniquely identify
+/// the anchored key regardless of the specific Value types involved.
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct AnchoredKeyId {
+    /// The dictionary root value (raw representation for Ord).
+    pub dict: RawValue,
+    /// The key within the dictionary (raw representation for Ord).
+    pub key: RawValue,
+}
+
+impl AnchoredKeyId {
+    /// Create a new anchored key ID from raw values.
+    pub fn new(dict: RawValue, key: RawValue) -> Self {
+        Self { dict, key }
+    }
+
+    /// Try to extract an anchored key ID from a Contains statement with all literal values.
+    pub fn from_contains_statement(stmt: &Statement) -> Option<Self> {
+        if let Statement::Contains(
+            ValueRef::Literal(dict),
+            ValueRef::Literal(key),
+            ValueRef::Literal(_value),
+        ) = stmt
+        {
+            Some(Self::new(dict.raw(), key.raw()))
+        } else {
+            None
+        }
+    }
+}
+
+/// Resource costs for a single statement/operation.
+///
+/// Each field corresponds to a resource with a per-POD limit in `Params`.
+#[derive(Clone, Debug, Default)]
+pub struct StatementCost {
+    /// Number of merkle proofs used (for Contains/NotContains).
+    /// Limit: `params.max_merkle_proofs_containers`
+    pub merkle_proofs: usize,
+
+    /// Number of merkle tree state transition proofs (for Insert/Update/Delete).
+    /// Limit: `params.max_merkle_tree_state_transition_proofs_containers`
+    pub merkle_state_transitions: usize,
+
+    /// Number of custom predicate verifications.
+    /// Limit: `params.max_custom_predicate_verifications`
+    pub custom_pred_verifications: usize,
+
+    /// Number of SignedBy operations.
+    /// Limit: `params.max_signed_by`
+    pub signed_by: usize,
+
+    /// Number of PublicKeyOf operations.
+    /// Limit: `params.max_public_key_of`
+    pub public_key_of: usize,
+
+    /// Custom predicate batches used (for batch cardinality constraint).
+    /// Limit: `params.max_custom_predicate_batches` distinct batches per POD.
+    pub custom_batch_ids: BTreeSet<CustomBatchId>,
+
+    /// Anchored keys referenced by this operation.
+    ///
+    /// When a Contains statement with all literal values is used as an argument,
+    /// the operation references an "anchored key" (dict, key pair). Each unique
+    /// anchored key used in a POD incurs an additional Contains statement cost,
+    /// as MainPodBuilder::add_entries_contains will auto-insert it if not already present.
+    pub anchored_keys: BTreeSet<AnchoredKeyId>,
+}
+
+impl StatementCost {
+    /// Compute the resource cost of an operation.
+    pub fn from_operation(op: &Operation) -> Self {
+        let mut cost = Self::default();
+
+        match &op.0 {
+            OperationType::Native(native_op) => {
+                match native_op {
+                    // Operations that use merkle proofs
+                    NativeOperation::ContainsFromEntries
+                    | NativeOperation::NotContainsFromEntries
+                    | NativeOperation::DictContainsFromEntries
+                    | NativeOperation::DictNotContainsFromEntries
+                    | NativeOperation::SetContainsFromEntries
+                    | NativeOperation::SetNotContainsFromEntries
+                    | NativeOperation::ArrayContainsFromEntries => {
+                        cost.merkle_proofs = 1;
+                    }
+
+                    // Operations that use merkle state transitions
+                    NativeOperation::ContainerInsertFromEntries
+                    | NativeOperation::ContainerUpdateFromEntries
+                    | NativeOperation::ContainerDeleteFromEntries
+                    | NativeOperation::DictInsertFromEntries
+                    | NativeOperation::DictUpdateFromEntries
+                    | NativeOperation::DictDeleteFromEntries
+                    | NativeOperation::SetInsertFromEntries
+                    | NativeOperation::SetDeleteFromEntries
+                    | NativeOperation::ArrayUpdateFromEntries => {
+                        cost.merkle_state_transitions = 1;
+                    }
+
+                    // SignedBy operation
+                    NativeOperation::SignedBy => {
+                        cost.signed_by = 1;
+                    }
+
+                    // PublicKeyOf operation
+                    NativeOperation::PublicKeyOf => {
+                        cost.public_key_of = 1;
+                    }
+
+                    // Operations with no special resource costs
+                    NativeOperation::None
+                    | NativeOperation::CopyStatement
+                    | NativeOperation::EqualFromEntries
+                    | NativeOperation::NotEqualFromEntries
+                    | NativeOperation::LtEqFromEntries
+                    | NativeOperation::LtFromEntries
+                    | NativeOperation::TransitiveEqualFromStatements
+                    | NativeOperation::LtToNotEqual
+                    | NativeOperation::SumOf
+                    | NativeOperation::ProductOf
+                    | NativeOperation::MaxOf
+                    | NativeOperation::HashOf
+                    // Syntactic sugar variants (lowered before proving)
+                    | NativeOperation::GtEqFromEntries
+                    | NativeOperation::GtFromEntries
+                    | NativeOperation::GtToNotEqual => {}
+                }
+            }
+            OperationType::Custom(cpr) => {
+                cost.custom_pred_verifications = 1;
+                cost.custom_batch_ids
+                    .insert(CustomBatchId::from(&*cpr.batch));
+            }
+        }
+
+        // Extract anchored keys from operation arguments.
+        // Any argument that is a Contains statement with all literal values
+        // represents an anchored key reference that will require a Contains
+        // statement in the POD (auto-inserted by MainPodBuilder if needed).
+        for arg in &op.1 {
+            if let OperationArg::Statement(stmt) = arg {
+                if let Some(anchored_key) = AnchoredKeyId::from_contains_statement(stmt) {
+                    cost.anchored_keys.insert(anchored_key);
+                }
+            }
+        }
+
+        cost
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::{
+        frontend::Operation as FrontendOp,
+        middleware::{NativeOperation, OperationAux, OperationType},
+    };
+
+    fn make_native_op(native_op: NativeOperation) -> FrontendOp {
+        FrontendOp(OperationType::Native(native_op), vec![], OperationAux::None)
+    }
+
+    #[test]
+    fn test_cost_from_native_ops() {
+        // Test merkle proof ops
+        let contains_op = make_native_op(NativeOperation::ContainsFromEntries);
+        let cost = StatementCost::from_operation(&contains_op);
+        assert_eq!(cost.merkle_proofs, 1);
+        assert_eq!(cost.merkle_state_transitions, 0);
+
+        // Test merkle state transition ops
+        let insert_op = make_native_op(NativeOperation::ContainerInsertFromEntries);
+        let cost = StatementCost::from_operation(&insert_op);
+        assert_eq!(cost.merkle_proofs, 0);
+        assert_eq!(cost.merkle_state_transitions, 1);
+
+        // Test signed_by
+        let signed_op = make_native_op(NativeOperation::SignedBy);
+        let cost = StatementCost::from_operation(&signed_op);
+        assert_eq!(cost.signed_by, 1);
+
+        // Test public_key_of
+        let pk_op = make_native_op(NativeOperation::PublicKeyOf);
+        let cost = StatementCost::from_operation(&pk_op);
+        assert_eq!(cost.public_key_of, 1);
+    }
+}

src/frontend/multi_pod/deps.rs

@@ -0,0 +1,178 @@
+//! Dependency analysis for statements and operations.
+//!
+//! This module analyzes dependencies between statements to determine
+//! which statements must be proved before others.
+
+use std::collections::HashMap;
+
+use crate::{
+    frontend::{Operation, OperationArg},
+    middleware::{Hash, Statement},
+};
+
+/// Represents a source of a statement dependency.
+#[derive(Clone, Debug, PartialEq, Eq, Hash)]
+pub enum StatementSource {
+    /// Statement created within this builder at the given index.
+    Internal(usize),
+    /// Statement from an external input POD (identified by POD hash).
+    External(Hash),
+}
+
+/// Dependency graph for all statements in a builder.
+///
+/// Each element `statement_deps[i]` is the list of dependencies for statement `i`.
+#[derive(Clone, Debug)]
+pub struct DependencyGraph {
+    /// Dependencies for each statement (indexed by statement index).
+    pub statement_deps: Vec<Vec<StatementSource>>,
+}
+
+impl DependencyGraph {
+    /// Build a dependency graph from statements and operations.
+    ///
+    /// `statements` and `operations` should be parallel arrays where
+    /// `operations[i]` produces `statements[i]`.
+    ///
+    /// `external_pod_statements` maps (pod_hash, statement) pairs to enable
+    /// recognizing references to external POD statements.
+    pub fn build(
+        statements: &[Statement],
+        operations: &[Operation],
+        external_pod_statements: &HashMap<Statement, Hash>,
+    ) -> Self {
+        let mut statement_deps = Vec::with_capacity(statements.len());
+
+        // Build a map from statement to its index for internal lookup.
+        // Use entry().or_insert() to preserve the FIRST occurrence of each statement.
+        // This is important for CopyStatement: if statements[0] = A and statements[2] = copy(A) = A,
+        // we want statement_to_index[A] = 0 (the original), not 2 (the copy).
+        let mut statement_to_index: HashMap<&Statement, usize> = HashMap::new();
+        for (i, s) in statements.iter().enumerate() {
+            if !s.is_none() {
+                statement_to_index.entry(s).or_insert(i);
+            }
+        }
+
+        for (idx, op) in operations.iter().enumerate() {
+            let mut deps = Vec::new();
+
+            // Examine each argument to the operation
+            for arg in &op.1 {
+                if let OperationArg::Statement(ref dep_stmt) = arg {
+                    if dep_stmt.is_none() {
+                        continue;
+                    }
+
+                    // Check if this is an internal statement (created earlier in this builder)
+                    if let Some(&dep_idx) = statement_to_index.get(dep_stmt) {
+                        // Internal dependencies must always be from earlier statements
+                        assert!(
+                            dep_idx <= idx,
+                            "Statement at index {} depends on future statement at index {}",
+                            idx,
+                            dep_idx
+                        );
+
+                        if dep_idx < idx {
+                            // The statement was created by an earlier operation
+                            deps.push(StatementSource::Internal(dep_idx));
+                            continue;
+                        }
+                        // dep_idx == idx: The first occurrence of this statement is at the current index,
+                        // meaning this operation both takes and produces this statement (e.g., CopyStatement
+                        // copying from an external POD). Fall through to check external PODs for the source.
+                    }
+
+                    // Check if this is from an external POD
+                    if let Some(&pod_hash) = external_pod_statements.get(dep_stmt) {
+                        deps.push(StatementSource::External(pod_hash));
+                    } else {
+                        // Statement arguments should either be internal (created earlier)
+                        // or from external PODs. If neither, something is wrong.
+                        unreachable!(
+                            "Statement argument not found in internal statements or external PODs: {:?}",
+                            dep_stmt
+                        );
+                    }
+                }
+            }
+
+            statement_deps.push(deps);
+        }
+
+        Self { statement_deps }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::{
+        frontend::Operation as FrontendOp,
+        middleware::{NativeOperation, OperationAux, OperationType, Value, ValueRef},
+    };
+
+    fn equal_stmt(n: i64) -> Statement {
+        Statement::Equal(
+            ValueRef::Literal(Value::from(n)),
+            ValueRef::Literal(Value::from(n)),
+        )
+    }
+
+    /// None operation produces Statement::None
+    fn none_op() -> FrontendOp {
+        FrontendOp(
+            OperationType::Native(NativeOperation::None),
+            vec![],
+            OperationAux::None,
+        )
+    }
+
+    /// CopyStatement(s) produces s (the same statement)
+    fn copy_op(stmt: Statement) -> FrontendOp {
+        FrontendOp(
+            OperationType::Native(NativeOperation::CopyStatement),
+            vec![OperationArg::Statement(stmt)],
+            OperationAux::None,
+        )
+    }
+
+    #[test]
+    fn test_copy_creates_dependency_on_original() {
+        // CopyStatement(s) produces s. When we copy a statement, the copy
+        // depends on where that statement first appears.
+        //
+        // statements[0] = s (produced by none_op - not realistic, but we need a first occurrence)
+        // statements[1] = s (produced by copy_op(s))
+        //
+        // op1's argument s matches statements[0], so statement 1 depends on statement 0.
+        let s = equal_stmt(1);
+
+        let statements = vec![s.clone(), s.clone()];
+        let operations = vec![
+            none_op(),  // Placeholder - in reality something else would produce s
+            copy_op(s), // Copies s, producing s. Depends on statements[0].
+        ];
+
+        let graph = DependencyGraph::build(&statements, &operations, &HashMap::new());
+
+        assert!(graph.statement_deps[0].is_empty());
+        assert_eq!(graph.statement_deps[1], vec![StatementSource::Internal(0)]);
+    }
+
+    #[test]
+    fn test_multiple_copies_depend_on_original() {
+        // Multiple copies of the same statement all depend on where it first appears.
+        let s = equal_stmt(1);
+
+        let statements = vec![s.clone(), s.clone(), s.clone()];
+        let operations = vec![none_op(), copy_op(s.clone()), copy_op(s)];
+
+        let graph = DependencyGraph::build(&statements, &operations, &HashMap::new());
+
+        assert!(graph.statement_deps[0].is_empty());
+        assert_eq!(graph.statement_deps[1], vec![StatementSource::Internal(0)]);
+        assert_eq!(graph.statement_deps[2], vec![StatementSource::Internal(0)]);
+    }
+}

src/frontend/multi_pod/solver.rs

@@ -0,0 +1,703 @@
+//! MILP solver for multi-POD packing.
+//!
+//! This module builds and solves a Mixed Integer Linear Program (MILP) to minimize
+//! the number of PODs needed to prove a set of statements while respecting resource
+//! limits and dependency constraints.
+//!
+//! # Constraint Overview
+//!
+//! The solver uses the following constraints (numbered for reference in code comments):
+//!
+//! - **Constraint 1 (Coverage)**: Each statement must be proved in at least one POD.
+//! - **Constraint 2 (Output POD)**: Output-public statements must be public in the last POD.
+//! - **Constraint 2b (Privacy)**: Non-output-public statements cannot be public in the output POD.
+//! - **Constraint 3 (Public ⇒ Proved)**: A statement can only be public if it's proved there.
+//! - **Constraint 4 (POD Existence)**: If any statement is proved in POD p, then p is used.
+//! - **Constraint 5 (Dependencies)**: If statement S depends on D and S is proved in POD p,
+//!   then D must be available: either proved locally in p, or public in some earlier POD.
+//! - **Constraint 5b (Copy Tracking)**: Track when dependencies need CopyStatement.
+//! - **Constraint 6 (Resource Limits)**: Per-POD limits on statements, public slots, merkle
+//!   proofs, custom predicates, batches, etc.
+//! - **Constraint 7 (Batch Cardinality)**: Limit distinct custom predicate batches per POD.
+//! - **Constraint 7b (Anchored Keys)**: Track auto-inserted Contains for anchored key references.
+//! - **Constraint 8a (Internal Inputs)**: Track which earlier PODs are used as inputs.
+//! - **Constraint 8b (External Inputs)**: Track which external PODs are used as inputs.
+//! - **Constraint 8c (Input Limit)**: Total inputs (internal + external) ≤ max_input_pods.
+//! - **Constraint 9 (Symmetry Breaking)**: PODs are used in order (0, 1, 2, ...) with no gaps.
+//!
+//! # Solution Approach
+//!
+//! The solver uses an incremental approach: it tries solving with the minimum possible
+//! number of PODs first, then increments until a feasible solution is found. This is
+//! efficient for the common case where few PODs are needed.
+
+// MILP constraint building uses explicit index loops for clarity
+#![allow(clippy::needless_range_loop)]
+
+use std::collections::BTreeSet;
+
+use good_lp::{
+    constraint, default_solver, variable, Expression, ProblemVariables, Solution, SolverModel,
+    Variable,
+};
+use itertools::Itertools;
+
+use super::Result;
+use crate::{
+    frontend::multi_pod::{
+        cost::{AnchoredKeyId, CustomBatchId, StatementCost},
+        deps::{DependencyGraph, StatementSource},
+    },
+    middleware::Params,
+};
+
+/// Threshold for interpreting MILP solver's floating-point results as binary.
+/// The solver returns continuous values in [0, 1] for binary variables;
+/// values > 0.5 are interpreted as "true" (1), otherwise "false" (0).
+const SOLVER_BINARY_THRESHOLD: f64 = 0.5;
+
+/// Solution from the MILP solver.
+#[derive(Clone, Debug)]
+pub struct MultiPodSolution {
+    /// Number of PODs needed.
+    pub pod_count: usize,
+
+    /// For each statement index, which POD(s) it is proved in.
+    /// (A statement may be proved in multiple PODs if re-proving is cheaper than copying.)
+    pub statement_to_pods: Vec<Vec<usize>>,
+
+    /// For each POD, which statement indices are proved in it.
+    pub pod_statements: Vec<Vec<usize>>,
+
+    /// For each POD, which statement indices are public in it.
+    pub pod_public_statements: Vec<BTreeSet<usize>>,
+}
+
+/// Input to the MILP solver.
+pub struct SolverInput<'a> {
+    /// Number of statements.
+    pub num_statements: usize,
+
+    /// Resource costs for each statement.
+    pub costs: &'a [StatementCost],
+
+    /// Dependency graph.
+    pub deps: &'a DependencyGraph,
+
+    /// Indices of statements that must be public in output PODs.
+    pub output_public_indices: &'a [usize],
+
+    /// Parameters defining per-POD limits.
+    pub params: &'a Params,
+
+    /// Maximum number of PODs the solver will consider.
+    pub max_pods: usize,
+
+    /// All unique anchored keys referenced by any statement.
+    ///
+    /// Each unique (dict, key) pair that is used as an anchored key reference
+    /// in any operation. When a Contains statement with literal values is used
+    /// as an argument, it creates an anchored key reference.
+    pub all_anchored_keys: &'a [AnchoredKeyId],
+
+    /// For each anchored key, the statement index that produces it (if any).
+    ///
+    /// When a Contains statement with literal (dict, key, value) args is explicitly
+    /// added, it "produces" that anchored key. If the producer is in the same POD
+    /// as statements using the anchored key, no auto-insertion is needed.
+    /// `anchored_key_producers[i]` corresponds to `all_anchored_keys[i]`.
+    pub anchored_key_producers: &'a [Option<usize>],
+
+    /// Statement content groups for deduplication.
+    ///
+    /// Each inner Vec contains statement indices that have identical content.
+    /// When multiple statements with the same content are proved in the same POD,
+    /// they only use one statement slot (the POD deduplicates identical statements).
+    pub statement_content_groups: &'a [Vec<usize>],
+}
+
+/// Solve the MILP problem to find optimal POD packing.
+///
+/// Uses an incremental approach: tries solving with min_pods first,
+/// then increments until a solution is found or target_pods is exceeded.
+/// This is efficient for the common case where min_pods is sufficient.
+pub fn solve(input: &SolverInput) -> Result<MultiPodSolution> {
+    let n = input.num_statements;
+
+    // Require at least one public statement. A POD with no public statements
+    // can't prove anything to an external verifier.
+    if input.output_public_indices.is_empty() {
+        return Err(super::Error::Solver(
+            "No public statements requested. Use pub_op() to add at least one statement \
+             that should be visible in the output POD."
+                .to_string(),
+        ));
+    }
+
+    // Check that all output-public statements can fit in a single POD
+    let num_output_public = input.output_public_indices.len();
+    if num_output_public > input.params.max_public_statements {
+        return Err(super::Error::Solver(format!(
+            "Too many public statements requested: {} requested, but max_public_statements is {}. \
+             All public statements must fit in a single output POD.",
+            num_output_public, input.params.max_public_statements
+        )));
+    }
+
+    // Lower bound on number of PODs needed
+    // Note: max_priv_statements is the limit on total unique statements per POD
+    // (public statements are copies from private slots)
+    let max_stmts_per_pod = input.params.max_priv_statements();
+    let min_pods_by_statements = n.div_ceil(max_stmts_per_pod);
+    let min_pods = min_pods_by_statements.max(1);
+
+    // Check if the problem exceeds the configured max_pods limit
+    if min_pods > input.max_pods {
+        return Err(super::Error::Solver(format!(
+            "Problem requires at least {} PODs, but max_pods is set to {}. \
+             Increase Options::max_pods to allow more PODs.",
+            min_pods, input.max_pods
+        )));
+    }
+
+    // Collect all unique custom batch IDs used
+    let all_batches: Vec<CustomBatchId> = input
+        .costs
+        .iter()
+        .flat_map(|c| c.custom_batch_ids.iter().cloned())
+        .unique()
+        .collect();
+
+    // Incremental approach: try solving with increasing POD counts
+    // Start with min_pods and increment until we find a feasible solution
+    for target_pods in min_pods..=input.max_pods {
+        if let Some(solution) = try_solve_with_pods(input, target_pods, &all_batches)? {
+            return Ok(solution);
+        }
+        // Infeasible with target_pods, try more
+    }
+
+    // No feasible solution found even with max_pods
+    Err(super::Error::Solver(format!(
+        "No feasible solution found with up to {} PODs",
+        input.max_pods
+    )))
+}
+
+/// Try to solve the packing problem with exactly `target_pods` PODs.
+///
+/// Builds a MILP model with all constraints and attempts to solve it.
+/// Returns `Ok(Some(solution))` if a feasible assignment exists,
+/// `Ok(None)` if the problem is infeasible with this many PODs.
+///
+/// The caller (in `solve()`) handles incrementing `target_pods` when infeasible.
+fn try_solve_with_pods(
+    input: &SolverInput,
+    target_pods: usize,
+    all_batches: &[CustomBatchId],
+) -> Result<Option<MultiPodSolution>> {
+    // Create variables
+    let mut vars = ProblemVariables::new();
+    let n = input.num_statements;
+
+    // prove[s][p] - statement s is proved in POD p
+    let prove: Vec<Vec<Variable>> = (0..n)
+        .map(|_| {
+            (0..target_pods)
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // public[s][p] - statement s is public in POD p
+    let public: Vec<Vec<Variable>> = (0..n)
+        .map(|_| {
+            (0..target_pods)
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // pod_used[p] - POD p is used
+    let pod_used: Vec<Variable> = (0..target_pods)
+        .map(|_| vars.add(variable().binary()))
+        .collect();
+
+    // batch_used[b][p] - custom batch b is used in POD p
+    let batch_used: Vec<Vec<Variable>> = (0..all_batches.len())
+        .map(|_| {
+            (0..target_pods)
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // anchored_key_used[ak][p] - anchored key ak is used in POD p
+    // When a statement references an anchored key (via a Contains statement argument),
+    // that POD must have a Contains statement for that (dict, key) pair.
+    // MainPodBuilder::add_entries_contains auto-inserts these, and we must account
+    // for them in the statement count.
+    let anchored_key_used: Vec<Vec<Variable>> = (0..input.all_anchored_keys.len())
+        .map(|_| {
+            (0..target_pods)
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // uses_input[p][pp] - POD p uses POD pp as an input (pp < p)
+    // We only create variables for pp < p
+    let uses_input: Vec<Vec<Variable>> = (0..target_pods)
+        .map(|p| (0..p).map(|_| vars.add(variable().binary())).collect())
+        .collect();
+
+    // Collect all statement indices that are internal dependencies.
+    // These are statements that other statements depend on, and may need to be copied
+    // into PODs where the dependent statement is proved but the dependency is not.
+    let internal_deps: BTreeSet<usize> = input
+        .deps
+        .statement_deps
+        .iter()
+        .flat_map(|deps| deps.iter())
+        .filter_map(|dep| match dep {
+            StatementSource::Internal(d) => Some(*d),
+            StatementSource::External(_) => None,
+        })
+        .collect();
+
+    // needs_copy[d][p] - dependency d needs to be copied into POD p
+    // This is 1 when: (some statement s in p depends on d) AND (d is not proved in p)
+    // We only create variables for dependencies that are actually used.
+    let dep_indices: Vec<usize> = internal_deps.iter().copied().collect();
+    let needs_copy: Vec<Vec<Variable>> = (0..dep_indices.len())
+        .map(|_| {
+            (0..target_pods)
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // Collect all external POD hashes that statements depend on.
+    // These are user-provided input PODs referenced by statements.
+    use crate::middleware::Hash;
+    let external_pods: Vec<Hash> = input
+        .deps
+        .statement_deps
+        .iter()
+        .flat_map(|deps| deps.iter())
+        .filter_map(|dep| match dep {
+            StatementSource::External(h) => Some(*h),
+            StatementSource::Internal(_) => None,
+        })
+        .collect::<BTreeSet<_>>()
+        .into_iter()
+        .collect();
+
+    // uses_external[p][e] - POD p uses external POD e as an input
+    let uses_external: Vec<Vec<Variable>> = (0..target_pods)
+        .map(|_| {
+            (0..external_pods.len())
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // Map from external POD hash to index in uses_external
+    let external_to_idx: std::collections::HashMap<Hash, usize> = external_pods
+        .iter()
+        .enumerate()
+        .map(|(i, h)| (*h, i))
+        .collect();
+
+    // content_group_used[g][p] - content group g has at least one statement proved in POD p
+    // When multiple statements have identical content, they share a slot in the POD.
+    // This variable tracks whether at least one statement from each content group is proved.
+    let num_groups = input.statement_content_groups.len();
+    let content_group_used: Vec<Vec<Variable>> = (0..num_groups)
+        .map(|_| {
+            (0..target_pods)
+                .map(|_| vars.add(variable().binary()))
+                .collect()
+        })
+        .collect();
+
+    // Objective: minimize number of PODs used
+    let objective: Expression = pod_used.iter().sum();
+    let mut model = vars.minimise(objective).using(default_solver);
+
+    // Constraint 1: Each statement must be proved at least once
+    for s in 0..n {
+        let sum: Expression = prove[s].iter().sum();
+        model.add_constraint(constraint!(sum >= 1));
+    }
+
+    // Constraint 2: Output-public statements must be public in the output POD (last POD)
+    // The output POD is at index target_pods-1, allowing it to access all earlier PODs
+    // for dependencies. This ensures exactly one output POD with deterministic location.
+    let output_pod = target_pods - 1;
+    for &s in input.output_public_indices {
+        model.add_constraint(constraint!(public[s][output_pod] == 1));
+    }
+
+    // Constraint 2b: Non-output-public statements cannot be public in the output POD
+    // This prevents private statements from leaking to the output POD's public slots.
+    for s in 0..n {
+        if !input.output_public_indices.contains(&s) {
+            model.add_constraint(constraint!(public[s][output_pod] == 0));
+        }
+    }
+
+    // Constraint 3: Public implies proved
+    for s in 0..n {
+        for p in 0..target_pods {
+            model.add_constraint(constraint!(public[s][p] <= prove[s][p]));
+        }
+    }
+
+    // Constraint 4: Pod existence - if any statement is proved in p, p is used
+    for s in 0..n {
+        for p in 0..target_pods {
+            model.add_constraint(constraint!(prove[s][p] <= pod_used[p]));
+        }
+    }
+
+    // Constraint 5: Dependencies (works with Constraint 8 to enforce input POD tracking)
+    //
+    // If s depends on d (internal), and s is proved in p, then either:
+    // - d is proved in p (local availability), OR
+    // - d is public in some earlier POD p' < p (cross-POD availability)
+    //
+    // This constraint ensures dependencies are AVAILABLE. It does NOT track which
+    // earlier PODs are actually used as inputs - that's handled by Constraint 8.
+    // Together:
+    // - Constraint 5 ensures the dependency CAN be satisfied
+    // - Constraint 8 ensures that when we use a statement from earlier POD pp,
+    //   we count pp as an input to pod p (for max_input_pods enforcement)
+    for s in 0..n {
+        for dep in &input.deps.statement_deps[s] {
+            if let StatementSource::Internal(d) = dep {
+                for p in 0..target_pods {
+                    // prove[s][p] <= prove[d][p] + sum_{p' < p} public[d][p']
+                    let mut rhs: Expression = prove[*d][p].into();
+                    for pp in 0..p {
+                        rhs += public[*d][pp];
+                    }
+                    model.add_constraint(constraint!(prove[s][p] <= rhs));
+                }
+            }
+        }
+    }
+
+    // Constraint 5b: needs_copy tracking for cross-POD dependencies
+    // needs_copy[d][p] = 1 when: some statement s proved in p depends on d, AND d is not proved in p.
+    // This tracks CopyStatements that will be added during build_single_pod.
+    for (di, &d) in dep_indices.iter().enumerate() {
+        for p in 0..target_pods {
+            // needs_copy[d][p] >= prove[s][p] - prove[d][p] for each s that depends on d
+            // If s is in p (prove[s][p]=1) and d is not in p (prove[d][p]=0), then needs_copy >= 1
+            for s in 0..n {
+                let depends_on_d = input.deps.statement_deps[s]
+                    .iter()
+                    .any(|dep| matches!(dep, StatementSource::Internal(dep_d) if *dep_d == d));
+                if depends_on_d {
+                    model.add_constraint(constraint!(
+                        needs_copy[di][p] >= prove[s][p] - prove[d][p]
+                    ));
+                }
+            }
+
+            // needs_copy[d][p] <= 1 - prove[d][p]
+            // If d is proved locally (prove[d][p]=1), no copy needed (needs_copy <= 0)
+            model.add_constraint(constraint!(needs_copy[di][p] <= 1 - prove[d][p]));
+        }
+    }
+
+    // Constraint 6: Resource limits per POD
+    //
+    // 6a-pre: Content group tracking for statement deduplication
+    // When multiple statement indices have identical content, they share a single slot in the POD.
+    // content_group_used[g][p] = 1 iff at least one statement from group g is proved in POD p.
+    for (g, group) in input.statement_content_groups.iter().enumerate() {
+        for p in 0..target_pods {
+            // Lower bound: if any statement in the group is proved, the group is used
+            for &s in group {
+                model.add_constraint(constraint!(content_group_used[g][p] >= prove[s][p]));
+            }
+            // Upper bound: if no statements in the group are proved, the group is not used
+            let group_prove_sum: Expression = group.iter().map(|&s| prove[s][p]).sum();
+            model.add_constraint(constraint!(content_group_used[g][p] <= group_prove_sum));
+        }
+    }
+
+    for p in 0..target_pods {
+        // 6a: Unique statement count (unique content groups + CopyStatements + anchored key Contains)
+        // Statements with identical content share a slot, so we count content groups, not indices.
+        // CopyStatements and anchored key Contains also use statement slots.
+        // The total must not exceed max_priv_statements (= max_statements - max_public_statements).
+        let unique_stmt_sum: Expression = (0..num_groups).map(|g| content_group_used[g][p]).sum();
+        let copy_sum: Expression = (0..dep_indices.len()).map(|di| needs_copy[di][p]).sum();
+        let anchored_key_sum: Expression = (0..input.all_anchored_keys.len())
+            .map(|ak| anchored_key_used[ak][p])
+            .sum();
+        model.add_constraint(constraint!(
+            unique_stmt_sum + copy_sum + anchored_key_sum
+                <= (input.params.max_priv_statements() as f64) * pod_used[p]
+        ));
+
+        // 6b: Public statement count
+        let pub_sum: Expression = (0..n).map(|s| public[s][p]).sum();
+        model.add_constraint(constraint!(
+            pub_sum <= (input.params.max_public_statements as f64) * pod_used[p]
+        ));
+
+        // 6c: Merkle proofs
+        let merkle_sum: Expression = (0..n)
+            .map(|s| (input.costs[s].merkle_proofs as f64) * prove[s][p])
+            .sum();
+        model.add_constraint(constraint!(
+            merkle_sum <= (input.params.max_merkle_proofs_containers as f64) * pod_used[p]
+        ));
+
+        // 6d: Merkle state transitions
+        let mst_sum: Expression = (0..n)
+            .map(|s| (input.costs[s].merkle_state_transitions as f64) * prove[s][p])
+            .sum();
+        model.add_constraint(constraint!(
+            mst_sum
+                <= (input
+                    .params
+                    .max_merkle_tree_state_transition_proofs_containers as f64)
+                    * pod_used[p]
+        ));
+
+        // 6e: Custom predicate verifications
+        let cpv_sum: Expression = (0..n)
+            .map(|s| (input.costs[s].custom_pred_verifications as f64) * prove[s][p])
+            .sum();
+        model.add_constraint(constraint!(
+            cpv_sum <= (input.params.max_custom_predicate_verifications as f64) * pod_used[p]
+        ));
+
+        // 6f: SignedBy
+        let sb_sum: Expression = (0..n)
+            .map(|s| (input.costs[s].signed_by as f64) * prove[s][p])
+            .sum();
+        model.add_constraint(constraint!(
+            sb_sum <= (input.params.max_signed_by as f64) * pod_used[p]
+        ));
+
+        // 6g: PublicKeyOf
+        let pko_sum: Expression = (0..n)
+            .map(|s| (input.costs[s].public_key_of as f64) * prove[s][p])
+            .sum();
+        model.add_constraint(constraint!(
+            pko_sum <= (input.params.max_public_key_of as f64) * pod_used[p]
+        ));
+    }
+
+    // Constraint 7: Batch cardinality
+    // batch_used[b][p] >= prove[s][p] for all s that use batch b (batch is used if any statement uses it)
+    // batch_used[b][p] <= sum of prove[s][p] for all s using batch b (batch is 0 if no statements use it)
+    for (b, batch_id) in all_batches.iter().enumerate() {
+        for p in 0..target_pods {
+            let mut sum: Expression = 0.into();
+            for s in 0..n {
+                if input.costs[s].custom_batch_ids.contains(batch_id) {
+                    model.add_constraint(constraint!(batch_used[b][p] >= prove[s][p]));
+                    sum += prove[s][p];
+                }
+            }
+            model.add_constraint(constraint!(batch_used[b][p] <= sum));
+        }
+    }
+
+    // Batch count per POD
+    for p in 0..target_pods {
+        let batch_sum: Expression = (0..all_batches.len()).map(|b| batch_used[b][p]).sum();
+        model.add_constraint(constraint!(
+            batch_sum <= (input.params.max_custom_predicate_batches as f64) * pod_used[p]
+        ));
+    }
+
+    // Constraint 7b: Anchored key tracking
+    //
+    // anchored_key_used[ak][p] = 1 when auto-insertion of a Contains is needed for anchored key ak in POD p.
+    // This happens when: some statement using ak is in POD p, AND the producing Contains is NOT in POD p.
+    //
+    // If a Contains statement explicitly produces ak (anchored_key_producers[ak] = Some(prod_idx)):
+    //   - Lower: anchored_key_used[ak][p] >= prove[s][p] - prove[prod_idx][p] for all s using ak
+    //   - Upper: anchored_key_used[ak][p] <= 1 - prove[prod_idx][p]
+    //   This ensures overhead is 0 when the producer is in the same POD.
+    //
+    // If no Contains produces ak (anchored_key_producers[ak] = None):
+    //   - Lower: anchored_key_used[ak][p] >= prove[s][p] for all s using ak
+    //   - Upper: anchored_key_used[ak][p] <= sum of prove[s][p] for all s using ak
+    //   Auto-insertion is always needed when any user is present.
+    for (ak_idx, ak) in input.all_anchored_keys.iter().enumerate() {
+        let producer = input.anchored_key_producers[ak_idx];
+
+        for p in 0..target_pods {
+            let mut user_sum: Expression = 0.into();
+            for s in 0..n {
+                if input.costs[s].anchored_keys.contains(ak) {
+                    if let Some(prod_idx) = producer {
+                        // Producer exists: only count overhead if producer not in this POD
+                        model.add_constraint(constraint!(
+                            anchored_key_used[ak_idx][p] >= prove[s][p] - prove[prod_idx][p]
+                        ));
+                    } else {
+                        // No producer: always need auto-insertion if user is present
+                        model.add_constraint(constraint!(
+                            anchored_key_used[ak_idx][p] >= prove[s][p]
+                        ));
+                    }
+                    user_sum += prove[s][p];
+                }
+            }
+
+            if let Some(prod_idx) = producer {
+                // If producer is in POD, no auto-insertion needed (overhead = 0)
+                model.add_constraint(constraint!(
+                    anchored_key_used[ak_idx][p] <= 1 - prove[prod_idx][p]
+                ));
+            } else {
+                // No producer: overhead is bounded by whether any user is present
+                model.add_constraint(constraint!(anchored_key_used[ak_idx][p] <= user_sum));
+            }
+        }
+    }
+
+    // Constraint 8a: Internal input POD tracking using uses_input
+    // uses_input[p][pp] >= prove[s][p] + public[d][pp] - prove[d][p] - 1
+    // for each dependency (s depends on d)
+    //
+    // If s is proved in p and d is public in pp, we need pp as input UNLESS d is also
+    // proved locally in p. Subtracting prove[d][p] ensures that when d is re-proved
+    // locally (prove[d][p] = 1), the constraint becomes uses_input >= 0, which is
+    // always satisfied without forcing the input relationship.
+    for s in 0..n {
+        for dep in &input.deps.statement_deps[s] {
+            if let StatementSource::Internal(d) = dep {
+                for p in 1..target_pods {
+                    for pp in 0..p {
+                        model.add_constraint(constraint!(
+                            uses_input[p][pp] >= prove[s][p] + public[*d][pp] - prove[*d][p] - 1.0
+                        ));
+                    }
+                }
+            }
+        }
+    }
+
+    // Constraint 8b: External input POD tracking using uses_external
+    // If statement s is proved in POD p and s depends on external POD e, then uses_external[p][e] = 1
+    for s in 0..n {
+        for dep in &input.deps.statement_deps[s] {
+            if let StatementSource::External(h) = dep {
+                if let Some(&e) = external_to_idx.get(h) {
+                    for p in 0..target_pods {
+                        // If s is proved in p, then uses_external[p][e] = 1
+                        model.add_constraint(constraint!(uses_external[p][e] >= prove[s][p]));
+                    }
+                }
+            }
+        }
+    }
+
+    // Constraint 8c: Total input PODs (internal + external) must not exceed max_input_pods
+    // For each POD p, the total number of inputs is:
+    // - Internal inputs: PODs pp < p that provide public statements used by p
+    // - External inputs: User-provided PODs referenced by statements in p
+    for p in 0..target_pods {
+        let internal_sum: Expression = if p > 0 {
+            (0..p).map(|pp| uses_input[p][pp]).sum()
+        } else {
+            0.into()
+        };
+        let external_sum: Expression = (0..external_pods.len()).map(|e| uses_external[p][e]).sum();
+        model.add_constraint(constraint!(
+            internal_sum + external_sum <= (input.params.max_input_pods as f64) * pod_used[p]
+        ));
+    }
+
+    // Constraint 9: Symmetry breaking - use PODs in order
+    // pod_used[p] >= pod_used[p+1]
+    for p in 0..target_pods - 1 {
+        model.add_constraint(constraint!(pod_used[p] >= pod_used[p + 1]));
+    }
+
+    // Solve
+    let solution = match model.solve() {
+        Ok(sol) => sol,
+        Err(_) => {
+            // Infeasible with this number of PODs, try more
+            return Ok(None);
+        }
+    };
+
+    // Extract solution: count how many PODs are used.
+    // Symmetry breaking (Constraint 9) ensures PODs are used in order with no gaps.
+    let mut pod_count = 0;
+    for p in 0..target_pods {
+        if solution.value(pod_used[p]) > SOLVER_BINARY_THRESHOLD {
+            pod_count += 1;
+        }
+    }
+
+    let mut statement_to_pods: Vec<Vec<usize>> = vec![vec![]; n];
+    let mut pod_statements: Vec<Vec<usize>> = vec![vec![]; pod_count];
+    let mut pod_public_statements: Vec<BTreeSet<usize>> = vec![BTreeSet::new(); pod_count];
+
+    for s in 0..n {
+        for p in 0..pod_count {
+            if solution.value(prove[s][p]) > SOLVER_BINARY_THRESHOLD {
+                statement_to_pods[s].push(p);
+                pod_statements[p].push(s);
+            }
+            if solution.value(public[s][p]) > SOLVER_BINARY_THRESHOLD {
+                pod_public_statements[p].insert(s);
+            }
+        }
+    }
+
+    Ok(Some(MultiPodSolution {
+        pod_count,
+        statement_to_pods,
+        pod_statements,
+        pod_public_statements,
+    }))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_no_public_statements_error() {
+        // At least one public statement is required - otherwise the POD can't
+        // prove anything to an external verifier.
+        let params = Params::default();
+        let deps = DependencyGraph {
+            statement_deps: vec![],
+        };
+
+        let input = SolverInput {
+            num_statements: 0,
+            costs: &[],
+            deps: &deps,
+            output_public_indices: &[],
+            params: &params,
+            max_pods: 20,
+            all_anchored_keys: &[],
+            anchored_key_producers: &[],
+            statement_content_groups: &[],
+        };
+
+        let result = solve(&input);
+        assert!(result.is_err());
+        assert!(result
+            .unwrap_err()
+            .to_string()
+            .contains("No public statements requested"));
+    }
+}