diff --git a/pil-analyzer/src/dead_column_checker.rs b/pil-analyzer/src/dead_column_checker.rs
new file mode 100644
index 0000000000..a6e6b59b84
--- /dev/null
+++ b/pil-analyzer/src/dead_column_checker.rs
@@ -0,0 +1,111 @@
+use std::collections::HashSet;
+
+use powdr_ast::analyzed::{AlgebraicExpression, Analyzed, PolyID, PolynomialType, PublicDeclaration};
+use powdr_number::FieldElement;
+use powdr_parser_util::SourceRef;
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct DeadColumn {
+    pub name: String,
+    pub source: SourceRef,
+}
+
+/// Returns committed/witness columns (including array elements) that are declared but never used
+/// by any identity (constraints/lookups/permutations), nor by public declarations.
+pub fn dead_committed_columns<T: FieldElement>(analyzed: &Analyzed<T>) -> Vec<DeadColumn> {
+    let used_poly_ids = used_poly_ids(analyzed);
+
+    // Get all committed polynomials, filtering out auto-added symbols and non-committed polynomials
+    let committed_polynomials = analyzed
+        .definitions
+        .iter()
+        .filter(|(name, (sym, _))| {
+            !analyzed.auto_added_symbols.contains(*name)
+                && matches!(sym.kind, powdr_ast::analyzed::SymbolKind::Poly(PolynomialType::Committed))
+        });
+
+    // Filter all used polynomials from the committed polynomials
+    // Return vector of DeadColumn
+    committed_polynomials.flat_map(|(_name, (sym, _def))| {
+            sym.array_elements()
+                .map(|(elem_name, poly_id)| (elem_name, sym.source.clone(), poly_id))
+                .collect::<Vec<_>>()
+        })
+        .filter(|(_name, _src, poly_id)| !used_poly_ids.contains(poly_id))
+        .map(|(name, source, _poly_id)| DeadColumn { name, source })
+        .collect()
+}
+
+pub fn check_no_dead_committed_columns<T: FieldElement>(analyzed: &Analyzed<T>) -> Result<(), String> {
+    let dead = dead_committed_columns(analyzed);
+    if dead.is_empty() {
+        return Ok(());
+    }
+
+    let formatted = dead
+        .iter()
+        .map(|d| format!("- {} ({})", d.name, format_source(&d.source)))
+        .collect::<Vec<_>>()
+        .join("\n");
+
+    Err(format!(
+        "Dead committed columns detected (declared but never referenced by any constraint/lookup/permutation/public declaration):\n{formatted}"
+    ))
+}
+
+fn used_poly_ids<T: FieldElement>(analyzed: &Analyzed<T>) -> HashSet<PolyID> {
+    let mut used: HashSet<PolyID> = HashSet::new();
+
+    // Count usage from public declarations.
+    for decl in analyzed.public_declarations.values() {
+        used_from_public_declaration(decl, &mut used);
+    }
+
+    // Count usage from identities, after inlining intermediates.
+    for identity in analyzed.identities_with_inlined_intermediate_polynomials() {
+        for expr in identity
+            .left
+            .selector
+            .iter()
+            .chain(identity.left.expressions.iter())
+            .chain(identity.right.selector.iter())
+            .chain(identity.right.expressions.iter())
+        {
+            collect_poly_ids(expr, &mut used);
+        }
+    }
+
+    used
+}
+
+fn used_from_public_declaration(decl: &PublicDeclaration, used: &mut HashSet<PolyID>) {
+    // Note: poly_id is filled in during condensation.
+    if let Some(poly_id) = decl.polynomial.poly_id {
+        used.insert(poly_id);
+    }
+}
+
+fn collect_poly_ids<T>(expr: &AlgebraicExpression<T>, used: &mut HashSet<PolyID>) {
+    match expr {
+        AlgebraicExpression::Reference(r) => {
+            used.insert(r.poly_id);
+        }
+        AlgebraicExpression::BinaryOperation(op) => {
+            collect_poly_ids(&op.left, used);
+            collect_poly_ids(&op.right, used);
+        }
+        AlgebraicExpression::UnaryOperation(op) => {
+            collect_poly_ids(&op.expr, used);
+        }
+        AlgebraicExpression::PublicReference(_)
+        | AlgebraicExpression::Challenge(_)
+        | AlgebraicExpression::Number(_) => {}
+    }
+}
+
+fn format_source(source: &SourceRef) -> String {
+    let file = source.file_name.as_ref().map(|s| s.as_ref()).unwrap_or("<unknown>");
+    format!("{file}:{}..{}", source.start, source.end)
+}
+
+
diff --git a/pil-analyzer/src/lib.rs b/pil-analyzer/src/lib.rs
index 325a27721f..3f93aec375 100644
--- a/pil-analyzer/src/lib.rs
+++ b/pil-analyzer/src/lib.rs
@@ -2,6 +2,7 @@
 
 mod call_graph;
 mod condenser;
+mod dead_column_checker;
 pub mod evaluator;
 pub mod expression_processor;
 mod pil_analyzer;
@@ -24,6 +25,7 @@ use powdr_ast::{
 };
 
 pub use pil_analyzer::{analyze_ast, analyze_file, analyze_string};
+pub use dead_column_checker::{check_no_dead_committed_columns, dead_committed_columns, DeadColumn};
 
 pub trait AnalysisDriver: Clone + Copy {
     /// Turns a declaration into an absolute name.
diff --git a/pil-analyzer/src/pil_analyzer.rs b/pil-analyzer/src/pil_analyzer.rs
index 04e0eeafaa..f6c673a4e0 100644
--- a/pil-analyzer/src/pil_analyzer.rs
+++ b/pil-analyzer/src/pil_analyzer.rs
@@ -5,8 +5,10 @@ use std::iter::once;
 use std::path::{Path, PathBuf};
 use std::str::FromStr;
 
-use itertools::Itertools;
-use powdr_ast::parsed::asm::{parse_absolute_path, AbsoluteSymbolPath, SymbolPath};
+
+use powdr_ast::parsed::asm::{
+    parse_absolute_path, AbsoluteSymbolPath, SymbolPath,
+};
 use powdr_ast::parsed::types::Type;
 use powdr_ast::parsed::visitor::Children;
 use powdr_ast::parsed::{
@@ -21,7 +23,7 @@ use powdr_ast::analyzed::{
 use powdr_parser::parse_type;
 
 use crate::type_inference::{infer_types, ExpectedType};
-use crate::{side_effect_checker, AnalysisDriver};
+use crate::{AnalysisDriver, check_no_dead_committed_columns, side_effect_checker};
 
 use crate::statement_processor::{Counters, PILItem, StatementProcessor};
 use crate::{condenser, evaluator, expression_processor::ExpressionProcessor};
@@ -49,7 +51,11 @@ fn analyze<T: FieldElement>(files: Vec<PILFile>) -> Analyzed<T> {
     analyzer.process(files);
     analyzer.side_effect_check();
     analyzer.type_check();
-    analyzer.condense()
+    let analyzed = analyzer.condense();
+    if let Err(e) = check_no_dead_committed_columns(&analyzed) {
+        eprintln!("Error checking for dead committed columns: {e}");
+    }
+    analyzed
 }
 
 #[derive(Default)]
diff --git a/pil-analyzer/tests/dead_columns.rs b/pil-analyzer/tests/dead_columns.rs
new file mode 100644
index 0000000000..a648766355
--- /dev/null
+++ b/pil-analyzer/tests/dead_columns.rs
@@ -0,0 +1,93 @@
+use powdr_number::GoldilocksField;
+use powdr_pil_analyzer::{analyze_string, dead_committed_columns};
+
+#[test]
+fn dead_committed_columns_basic() {
+    let input = r#"
+        namespace N(16);
+        pol commit a;
+        pol commit b;
+        a = 1;
+    "#;
+
+    let analyzed = analyze_string::<GoldilocksField>(input);
+    let dead = dead_committed_columns(&analyzed);
+
+    assert!(dead.len() == 1, "expected 1 dead column, got: {dead:?}");
+    assert!(dead[0].name == "N.b", "expected N.b to be dead, got: {dead:?}");
+}
+
+
+#[test]
+fn dead_commited_test_1() {
+    let input = r#"
+        pol commit secret_value;
+    "#;
+    let analyzed = analyze_string::<GoldilocksField>(input);
+    let dead = dead_committed_columns(&analyzed);
+
+    assert!(dead.len() == 1, "expected 1 dead column, got: {dead:?}");
+    assert!(dead[0].name == "secret_value", "expected secret_value to be dead, got: {dead:?}");
+}
+
+#[test]
+fn dead_commited_test_2() {
+    let input = r#"
+        pol commit secret_value;
+        pol commit other_value;
+    "#;
+    let analyzed = analyze_string::<GoldilocksField>(input);
+    let dead = dead_committed_columns(&analyzed);
+    assert!(dead.len() == 2, "expected 1 dead column, got: {dead:?}");
+    assert!(dead[0].name == "secret_value", "expected secret_value to be dead, got: {dead:?}");
+    assert!(dead[1].name == "other_value", "expected other_value to be dead, got: {dead:?}");
+}
+
+#[test]
+fn used_in_itermidiate_then_constrained() {
+    let input = r#"
+        pol commit sel;
+        pol commit expected;
+        pol commit raw_value;
+        pol PROCESSED = raw_value * raw_value;
+        sel * (PROCESSED - expected) = 0;  // raw_value is constrained through PROCESSED
+    "#;
+    let analyzed = analyze_string::<GoldilocksField>(input);
+    let dead = dead_committed_columns(&analyzed);
+
+    assert!(dead.len() == 0, "expected no dead columns, got: {dead:?}");
+}
+
+#[test]
+fn used_in_intermediate_then_not_constrained() {
+    let input = r#"
+        pol commit raw_value;
+        pol PROCESSED = raw_value * raw_value;
+    "#;
+    let analyzed = analyze_string::<GoldilocksField>(input);
+    let dead = dead_committed_columns(&analyzed);
+
+    assert!(dead.len() == 1, "expected 1 dead column, got: {dead:?}");
+    assert!(dead[0].name == "raw_value", "expected raw_value to be dead, got: {dead:?}");
+}
+
+#[test]
+fn used_as_lookup_key_or_destination_is_not_dead() {
+    let input = r#"
+        namespace N(16);
+        pol commit sel;
+        pol commit query_value;
+        pol commit precomputed_value;
+
+        // If sel = 1, require query_value to appear in the "table" precomputed_value.
+        // This uses both committed columns only through the lookup identity.
+        sel { query_value } in precomputed_value { precomputed_value };
+    "#;
+    let analyzed = analyze_string::<GoldilocksField>(input);
+    let dead = dead_committed_columns(&analyzed);
+
+    assert!(
+        dead.len() == 0,
+        "expected no dead columns, got: {dead:?}"
+    );
+}