diff --git a/Cargo.lock b/Cargo.lock
index 7c50a92a1..113028ea1 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -5476,7 +5476,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb"
-version = "0.39.0-beta.5"
+version = "0.39.0-beta.6"
 dependencies = [
  "ahash",
  "anyhow",
@@ -5553,6 +5553,7 @@ dependencies = [
  "serde_json",
  "serde_with",
  "serial_test",
+ "sha2 0.10.9",
  "snafu 0.8.9",
  "tempfile",
  "test-log",
@@ -5567,7 +5568,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb-nodejs"
-version = "0.39.0-beta.5"
+version = "0.39.0-beta.6"
 dependencies = [
  "arrow-array",
  "arrow-buffer",
@@ -5592,7 +5593,7 @@ dependencies = [
 
 [[package]]
 name = "lancedb-python"
-version = "0.39.0-beta.5"
+version = "0.39.0-beta.6"
 dependencies = [
  "arrow",
  "async-trait",
diff --git a/docs/src/js/classes/Table.md b/docs/src/js/classes/Table.md
index 894d7a464..75e401d13 100644
--- a/docs/src/js/classes/Table.md
+++ b/docs/src/js/classes/Table.md
@@ -74,10 +74,10 @@ now: the column is committed with no values, and rows get them from
 [Table#refreshColumn](Table.md#refreshcolumn). Declaring one therefore costs the same on a
 large table as on an empty one.
 
-A refresh does not revisit rows it has already filled, so mutating an
-input leaves the value computed at fill time; recomputing means dropping
-the column and declaring it again. While a declaration reads a column,
-that column cannot be renamed, retyped or dropped.
+A refresh also recomputes the rows whose inputs changed since they were
+computed, so a mutated input is reflected by the next refresh. While a
+declaration reads a column, that column cannot be renamed, retyped or
+dropped.
 
 On LanceDB Cloud and Enterprise the expression is planned by the
 server, and the refresh runs as a server job -- see
@@ -854,10 +854,10 @@ abstract refreshColumn(column): Promise<RefreshColumnResult>
 
 Fill the rows of a computed column that hold no value yet.
 
-Rows appended since the last refresh are filled by the next one; rows
-already filled are left as they are, so the call is idempotent and does
-not observe a mutated input. Local tables only: a remote refresh runs
-as a server job, through [Table#refreshColumnAsync](Table.md#refreshcolumnasync).
+Rows appended since the last refresh are filled by the next one, and
+rows whose inputs changed since they were computed are recomputed;
+everything else is left as it is. Local tables only: a remote refresh
+runs as a server job, through [Table#refreshColumnAsync](Table.md#refreshcolumnasync).
 
 #### Parameters
 
diff --git a/nodejs/lancedb/table.ts b/nodejs/lancedb/table.ts
index 06f8cd991..34ea147c2 100644
--- a/nodejs/lancedb/table.ts
+++ b/nodejs/lancedb/table.ts
@@ -542,10 +542,10 @@ export abstract class Table {
    * {@link Table#refreshColumn}. Declaring one therefore costs the same on a
    * large table as on an empty one.
    *
-   * A refresh does not revisit rows it has already filled, so mutating an
-   * input leaves the value computed at fill time; recomputing means dropping
-   * the column and declaring it again. While a declaration reads a column,
-   * that column cannot be renamed, retyped or dropped.
+   * A refresh also recomputes the rows whose inputs changed since they were
+   * computed, so a mutated input is reflected by the next refresh. While a
+   * declaration reads a column, that column cannot be renamed, retyped or
+   * dropped.
    *
    * On LanceDB Cloud and Enterprise the expression is planned by the
    * server, and the refresh runs as a server job -- see
@@ -576,10 +576,10 @@ export abstract class Table {
   /**
    * Fill the rows of a computed column that hold no value yet.
    *
-   * Rows appended since the last refresh are filled by the next one; rows
-   * already filled are left as they are, so the call is idempotent and does
-   * not observe a mutated input. Local tables only: a remote refresh runs
-   * as a server job, through {@link Table#refreshColumnAsync}.
+   * Rows appended since the last refresh are filled by the next one, and
+   * rows whose inputs changed since they were computed are recomputed;
+   * everything else is left as it is. Local tables only: a remote refresh
+   * runs as a server job, through {@link Table#refreshColumnAsync}.
    * @param {string} column The name of the computed column to fill.
    * @returns {Promise<RefreshColumnResult>} A promise that resolves to the
    * number of rows filled and the new version number of the table.
diff --git a/python/python/lancedb/table.py b/python/python/lancedb/table.py
index 127ad3722..a41a4b8f9 100644
--- a/python/python/lancedb/table.py
+++ b/python/python/lancedb/table.py
@@ -2188,10 +2188,10 @@ class Table(ABC):
             Declaring one therefore costs the same on a large table as on an
             empty one.
 
-            A refresh does not revisit rows it has already filled, so mutating
-            an input leaves the value computed at fill time; recomputing means
-            dropping the column and declaring it again. While a declaration
-            reads a column, that column cannot be renamed, retyped or dropped.
+            A refresh also recomputes the rows whose inputs changed since they
+            were computed, so a mutated input is reflected by the next refresh.
+            While a declaration reads a column, that column cannot be renamed,
+            retyped or dropped.
 
             On LanceDB Cloud and Enterprise the expression is planned by the
             server, and the refresh runs as a server job -- see
@@ -2211,7 +2211,7 @@ class Table(ABC):
         >>> table.add_columns(computed={"doubled": "x * 2"})
         AddColumnsResult(version=2)
         >>> table.refresh_column("doubled")
-        RefreshColumnResult(rows_filled=2, version=3)
+        RefreshColumnResult(rows_filled=2, version=4)
         >>> table.to_arrow().sort_by("x").to_pandas()
            x  doubled
         0  1        2
@@ -2225,8 +2225,8 @@ class Table(ABC):
 
         Declared with ``add_columns(computed=...)``, a column starts empty and
         gets its values here. Rows appended since the last refresh are filled
-        by the next one; rows already filled are left as they are, so the call
-        is idempotent and does not observe a mutated input.
+        by the next one, and rows whose inputs changed since they were computed
+        are recomputed; everything else is left as it is.
 
         Local tables only: a remote refresh runs as a server job, through
         [`refresh_column_async`][lancedb.table.Table.refresh_column_async].
@@ -4318,13 +4318,14 @@ class LanceTable(Table):
         return LOOP.run(self._table.add_columns(transforms, computed=computed))
 
     def refresh_column(self, column: str) -> "RefreshColumnResult":
-        """Fill a computed column's unfilled rows. See
+        """Fill a computed column's unfilled rows and recompute those whose
+        inputs changed. See
         [`AsyncTable.refresh_column`][lancedb.AsyncTable.refresh_column]."""
         return LOOP.run(self._table.refresh_column(column))
 
     def refresh_column_async(self, column: str) -> Job[RefreshColumnJobResult]:
-        """Fill a computed column's unfilled rows, returning a handle to the
-        refresh job. See
+        """Fill a computed column's unfilled rows and recompute those whose
+        inputs changed, returning a handle to the refresh job. See
         [`Table.refresh_column_async`][lancedb.table.Table.refresh_column_async].
         """
         return Job(LOOP.run(self._table.refresh_column_async(column)))
@@ -6312,10 +6313,10 @@ class AsyncTable:
             them from
             [`refresh_column`][lancedb.table.AsyncTable.refresh_column].
 
-            A refresh does not revisit rows it has already filled, so mutating
-            an input leaves the value computed at fill time. While a
-            declaration reads a column, that column cannot be renamed, retyped
-            or dropped.
+            A refresh also recomputes the rows whose inputs changed since they
+            were computed, so a mutated input is reflected by the next refresh.
+            While a declaration reads a column, that column cannot be renamed,
+            retyped or dropped.
 
             On LanceDB Cloud and Enterprise the expression is planned by
             the server. Cannot be combined with ``transforms``.
@@ -6377,8 +6378,8 @@ class AsyncTable:
 
         Declared with ``add_columns(computed=...)``, a column starts empty and
         gets its values here. Rows appended since the last refresh are filled
-        by the next one; rows already filled are left as they are, so the call
-        is idempotent and does not observe a mutated input.
+        by the next one, and rows whose inputs changed since they were computed
+        are recomputed; everything else is left as it is.
 
         Local tables only: a remote refresh runs as a server job, through
         [`refresh_column_async`][lancedb.table.Table.refresh_column_async].
diff --git a/python/python/tests/test_table.py b/python/python/tests/test_table.py
index 3e1f6fb37..123219567 100644
--- a/python/python/tests/test_table.py
+++ b/python/python/tests/test_table.py
@@ -4183,13 +4183,14 @@ def test_refresh_column_async_returns_job(tmp_path):
     assert result.rows_failed == 0
     assert result.rows_remaining == 0
     assert result.source_version == 2
-    assert result.published_version == 3
+    # The fill lands at 3; the stamp recording its inputs is published at 4.
+    assert result.published_version == 4
     assert job.status() == "finished"
     assert sorted(table.to_arrow()["doubled"].to_pylist()) == [2, 4]
 
     no_op = table.refresh_column_async("doubled").wait()
     assert no_op.rows_assigned == 0
-    assert no_op.source_version == 3
+    assert no_op.source_version == 4
     assert no_op.published_version is None
 
     # Bad input raises at the call, not through the job.
@@ -4208,6 +4209,6 @@ async def test_refresh_column_async_job_async_table(tmp_path):
     assert isinstance(result, lancedb.RefreshColumnResult)
     assert result.rows_assigned == 1
     assert result.source_version == 2
-    assert result.published_version == 3
+    assert result.published_version == 4
     assert await job.status() == "finished"
     assert (await table.to_arrow())["tripled"].to_pylist() == [9]
diff --git a/rust/lancedb/Cargo.toml b/rust/lancedb/Cargo.toml
index c8f17d50c..188f6574c 100644
--- a/rust/lancedb/Cargo.toml
+++ b/rust/lancedb/Cargo.toml
@@ -95,13 +95,14 @@ candle-transformers = { version = "0.9.1", optional = true }
 candle-nn = { version = "0.9.1", optional = true }
 tokenizers = { version = "0.19.1", optional = true }
 semver = { workspace = true }
+roaring = "0.11.4"
+sha2 = "0.10"
 
 [dev-dependencies]
 anyhow = "1"
 lance-testing = { workspace = true }
 tempfile = { workspace = true }
 random_word = { version = "0.4.3", features = ["en"] }
-roaring = "0.11.4"
 tokio = { workspace = true, features = ["io-util", "macros", "net", "test-util"] }
 uuid = { workspace = true }
 walkdir = "2"
diff --git a/rust/lancedb/src/materialized_view/refresh.rs b/rust/lancedb/src/materialized_view/refresh.rs
index 23bb51566..433b185f7 100644
--- a/rust/lancedb/src/materialized_view/refresh.rs
+++ b/rust/lancedb/src/materialized_view/refresh.rs
@@ -1124,8 +1124,9 @@ struct RowScope {
 
 /// Whether every commit on the view after `recorded` is a fill of its
 /// computed columns: a column rewrite or data replacement touching only
-/// those fields and neither adding nor removing rows. A version whose
-/// transaction cannot be read is not proven, so it counts as drift.
+/// those fields and neither adding nor removing rows, or the freshness
+/// stamp a fill leaves on them. A version whose transaction cannot be read
+/// is not proven, so it counts as drift.
 async fn only_computed_rewrites_since(view_ds: &Dataset, recorded: u64) -> Result<bool> {
     // A fill may write any field under a computed column, so the whole
     // subtree counts, not only the root.
@@ -1176,6 +1177,19 @@ async fn only_computed_rewrites_since(view_ds: &Dataset, recorded: u64) -> Resul
                                 .all(|field| computed_fields.contains(&(*field as u32)))
                     })
             }
+            // The stamp `refresh_column` writes after its fill (see
+            // `table::freshness`): field metadata on computed columns, no data.
+            Operation::UpdateConfig {
+                config_updates: None,
+                table_metadata_updates: None,
+                schema_metadata_updates: None,
+                field_metadata_updates,
+            } => {
+                !field_metadata_updates.is_empty()
+                    && field_metadata_updates
+                        .keys()
+                        .all(|field| computed_fields.contains(&(*field as u32)))
+            }
             _ => false,
         };
         if !fill {
@@ -3359,6 +3373,46 @@ mod tests {
         );
     }
 
+    /// Field metadata on `field` only, the commit shape of the freshness
+    /// stamp `refresh_column` leaves after its fill.
+    async fn commit_field_metadata(view: &MaterializedView, field: &str, key: &str) {
+        let native = view.table().as_native().unwrap();
+        native.dataset.reload().await.unwrap();
+        let mut dataset = native.dataset.get().await.unwrap().as_ref().clone();
+        dataset
+            .update_field_metadata()
+            .update(field, [(key.to_string(), "{}".to_string())])
+            .unwrap()
+            .await
+            .unwrap();
+    }
+
+    /// The stamp is metadata on the computed column and rewrites nothing
+    /// refresh certifies, so it is not drift; the same commit shape on a
+    /// projected column is, like any other write to it.
+    #[tokio::test]
+    async fn test_a_freshness_stamp_is_not_drift() {
+        let conn = connect("memory://").execute().await.unwrap();
+        let view = refreshed_computed_view(&conn).await;
+
+        commit_field_metadata(
+            &view,
+            "emb",
+            crate::table::computed_columns::SOURCE_SIGNATURE_META_KEY,
+        )
+        .await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::NoOp
+        );
+
+        commit_field_metadata(&view, "id", "probe").await;
+        assert_eq!(
+            view.refresh().execute().await.unwrap().mode,
+            RefreshMode::Rebuild
+        );
+    }
+
     /// The fill job's commit rewrites only computed columns. It is the one
     /// commit on a view that is not drift: the next refresh carries on from
     /// its watermark instead of rebuilding, which would null what the fill
@@ -3524,9 +3578,9 @@ mod tests {
     }
 
     /// A SQL declaration is filled by `refresh_column` on the view, which
-    /// commits a data replacement; the next refresh continues from its
-    /// watermark and keeps what the fill wrote, and only rows the view added
-    /// since come back unfilled.
+    /// commits a data replacement and then its freshness stamp; the next
+    /// refresh continues from its watermark and keeps what the fill wrote,
+    /// and only rows the view added since come back unfilled.
     #[tokio::test]
     async fn test_a_sql_fill_is_not_drift() {
         use crate::materialized_view::tests::{people, sql_field};
diff --git a/rust/lancedb/src/table.rs b/rust/lancedb/src/table.rs
index 636aaefd0..3a334eaa2 100644
--- a/rust/lancedb/src/table.rs
+++ b/rust/lancedb/src/table.rs
@@ -75,6 +75,7 @@ mod create_index;
 pub mod datafusion;
 pub(crate) mod dataset;
 pub mod delete;
+pub mod freshness;
 pub mod lsm_stats;
 pub mod merge;
 pub mod optimize;
@@ -778,7 +779,8 @@ pub trait BaseTable: std::fmt::Display + std::fmt::Debug + Send + Sync {
             message: "Function columns are supported only on LanceDB Cloud and Enterprise".into(),
         })
     }
-    /// Fill a computed column's unfilled rows.
+    /// Fill a computed column's unfilled rows and recompute those whose
+    /// inputs changed.
     ///
     /// The default returns `NotSupported`; Lance-backed tables override it.
     async fn refresh_column(&self, _column: &str) -> Result<RefreshColumnResult> {
@@ -786,8 +788,8 @@ pub trait BaseTable: std::fmt::Display + std::fmt::Debug + Send + Sync {
             message: "computed columns are supported only on local tables".into(),
         })
     }
-    /// Fill a computed column's unfilled rows, returning a [`Job`] tracking
-    /// the operation.
+    /// Fill a computed column's unfilled rows and recompute those whose
+    /// inputs changed, returning a [`Job`] tracking the operation.
     async fn refresh_column_async(
         &self,
         _column: &str,
@@ -1749,9 +1751,10 @@ impl Table {
     /// Declared with
     /// [`AddColumnsBuilder::computed`](add_columns::AddColumnsBuilder::computed),
     /// a column starts empty and gets its values here. Fragments appended
-    /// since the last refresh are filled by the next one; fragments already
-    /// filled are left as they are, so the call is idempotent and does not
-    /// observe a mutated input.
+    /// since the last refresh are filled by the next one, and fragments whose
+    /// inputs changed since they were computed are recomputed (see
+    /// [`freshness`](crate::table::freshness)); everything else is left as
+    /// it is.
     ///
     /// Local tables only: a remote refresh runs as a server job, through
     /// [`Table::refresh_column_async`].
diff --git a/rust/lancedb/src/table/add_columns.rs b/rust/lancedb/src/table/add_columns.rs
index 1ac0c6b4f..635a99bea 100644
--- a/rust/lancedb/src/table/add_columns.rs
+++ b/rust/lancedb/src/table/add_columns.rs
@@ -60,10 +60,11 @@ impl AddColumnsBuilder {
     /// every fragment that has none -- including fragments appended since the
     /// last refresh.
     ///
-    /// Refresh does not revisit a fragment it has filled, so mutating an input
-    /// leaves the value computed at fill time; recomputing means dropping the
-    /// column and declaring it again. An input cannot be renamed, retyped or
-    /// dropped while a declaration reads it, since the expression names it.
+    /// A refresh also recomputes the rows of a fragment whose inputs changed
+    /// since it was computed (see [`freshness`](super::freshness)), so a
+    /// mutated input is reflected by the next refresh. An input cannot be
+    /// renamed, retyped or dropped while a declaration reads it, since the
+    /// expression names it.
     ///
     /// On LanceDB Cloud and Enterprise the expression is planned by the
     /// server, and the refresh runs as a server job -- see
diff --git a/rust/lancedb/src/table/computed_columns.rs b/rust/lancedb/src/table/computed_columns.rs
index f1ec75213..a70ac31ae 100644
--- a/rust/lancedb/src/table/computed_columns.rs
+++ b/rust/lancedb/src/table/computed_columns.rs
@@ -71,6 +71,22 @@ pub const FUNCTION_BINDINGS_META_KEY: &str = "lancedb::function_bindings";
 /// Version of the schema-level Function binding envelope.
 pub const FUNCTION_BINDINGS_VERSION: u32 = 1;
 
+/// Field metadata key holding `{fragment id -> input signature}` as JSON,
+/// recorded by the refresh that last computed each fragment. Outside the
+/// declaration namespace on purpose: a declaration is immutable through
+/// metadata edits, this is rewritten by every refresh. Seeded empty at
+/// declaration, so a column is tracked from birth; a column without it was
+/// declared before signatures existed.
+pub const SOURCE_SIGNATURE_META_KEY: &str = "computed_refresh.source_signature";
+
+/// Field metadata key holding the definition digest a column was last
+/// computed under. A change to it makes every row stale.
+pub const DEFINITION_VERSION_META_KEY: &str = "computed_refresh.definition_version";
+
+/// Field metadata key holding the table version the signature map describes:
+/// where a refresh starts following compactions to carry freshness forward.
+pub const RECORDED_AT_VERSION_META_KEY: &str = "computed_refresh.recorded_at_version";
+
 /// Value of [`KIND_META_KEY`] for a column defined by a SQL expression.
 pub const SQL_KIND: &str = "sql";
 
@@ -139,6 +155,7 @@ fn computed_column_metadata(expression: &str, inputs: &[String]) -> HashMap<Stri
             INPUTS_META_KEY.to_string(),
             serde_json::to_string(inputs).unwrap_or_else(|_| "[]".to_string()),
         ),
+        (SOURCE_SIGNATURE_META_KEY.to_string(), "{}".to_string()),
     ])
 }
 
@@ -163,6 +180,7 @@ pub fn function_computed_column_metadata(
             INPUTS_META_KEY.to_string(),
             serde_json::to_string(inputs).unwrap_or_else(|_| "[]".to_string()),
         ),
+        (SOURCE_SIGNATURE_META_KEY.to_string(), "{}".to_string()),
     ])
 }
 
@@ -1295,7 +1313,8 @@ pub(crate) fn ensure_not_an_input(schema: &SchemaRef, paths: &[&str]) -> Result<
 }
 
 /// Reject a write that supplies values for a computed column directly:
-/// only refresh materializes one, and refresh never revisits a filled row.
+/// only refresh materializes one, and only refresh decides what it
+/// recomputes.
 pub(crate) fn ensure_not_written<'a>(
     schema: &ArrowSchema,
     written: impl IntoIterator<Item = &'a str>,
@@ -1470,7 +1489,9 @@ fn ensure_no_foreign_declaration(field: &ArrowField) -> Result<()> {
 /// kind, the expression, the inputs -- would bypass that validation or move
 /// a binding out from under a refresh. Drop the column and declare it again.
 pub(crate) fn is_declaration_key(key: &str) -> bool {
-    key == COMPUTED_COLUMN_META_KEY || key.starts_with("computed_column.")
+    key == COMPUTED_COLUMN_META_KEY
+        || key.starts_with("computed_column.")
+        || key.starts_with("computed_refresh.")
 }
 
 /// Reject retyping a computed column itself.
diff --git a/rust/lancedb/src/table/freshness.rs b/rust/lancedb/src/table/freshness.rs
new file mode 100644
index 000000000..3295a0e7e
--- /dev/null
+++ b/rust/lancedb/src/table/freshness.rs
@@ -0,0 +1,1165 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+//! Source-change detection for computed columns.
+//!
+//! A refresh fills nulls, so once a value is durable nothing recomputes it and
+//! a later write to one of its inputs leaves it stale forever. Two stamps in
+//! the column's field metadata close that: the definition it was computed
+//! under, and a per-fragment signature of the input storage it was computed
+//! from. A refresh compares both with the manifest and recomputes what
+//! disagrees. Signatures are read from manifests, never from data.
+
+use std::collections::{BTreeMap, HashSet};
+use std::ops::Range;
+
+use arrow_array::{Array, UInt64Array};
+use futures::TryStreamExt;
+use lance::Dataset;
+use lance::dataset::transaction::Operation;
+use lance_core::ROW_ADDR;
+use lance_core::datatypes::{Field as LanceField, Schema as LanceSchema};
+use lance_table::format::{DataFile, Fragment};
+use roaring::RoaringBitmap;
+use serde::{Deserialize, Serialize};
+use sha2::{Digest, Sha256};
+
+use crate::table::computed_columns::{
+    DEFINITION_VERSION_META_KEY, RECORDED_AT_VERSION_META_KEY, SOURCE_SIGNATURE_META_KEY,
+};
+use crate::{Error, Result};
+
+/// A map recorded further back than this carries nothing through the
+/// compactions since: the fragments they produced are recomputed instead.
+const MAX_CARRY_FORWARD_VERSIONS: u64 = 1024;
+
+/// `{fragment id -> input signature}`.
+pub type SignatureMap = BTreeMap<u32, String>;
+
+/// Every field an input covers, keyed by column path: the field's own id
+/// first, then its ancestors', because a packed file records the physical
+/// column under an ancestor's id.
+pub type InputFields = BTreeMap<Vec<String>, Vec<i32>>;
+
+fn invalid(message: String) -> Error {
+    Error::InvalidInput { message }
+}
+
+/// FNV-1a over the text, as hex. Stable across processes and versions, which
+/// a signature compared against a stored one has to be.
+fn short_hash(value: &str) -> String {
+    let digest = Sha256::digest(value.as_bytes());
+    digest.iter().take(8).map(|b| format!("{b:02x}")).collect()
+}
+
+/// Digest of the definition a column is computed under.
+pub fn definition_version(definition: &str) -> String {
+    short_hash(definition)
+}
+
+/// The fields the named input column paths cover, children included. Paths,
+/// not ids, pair one schema with another: a rewrite renumbers fields. The key
+/// is the path's components, so a field named `a.b` and a nested `a` -> `b`
+/// are different columns.
+pub fn fields_for_paths(schema: &LanceSchema, paths: &[String]) -> Result<InputFields> {
+    fn collect(field: &LanceField, path: Vec<String>, ancestors: &[i32], out: &mut InputFields) {
+        let mut ids = vec![field.id];
+        ids.extend_from_slice(ancestors);
+        for child in &field.children {
+            let mut child_path = path.clone();
+            child_path.push(child.name.clone());
+            collect(child, child_path, &ids, out);
+        }
+        out.insert(path, ids);
+    }
+    let mut out = InputFields::new();
+    for field_path in paths {
+        let parts = lance_core::datatypes::parse_field_path(field_path)?;
+        let (root, rest) = parts
+            .split_first()
+            .ok_or_else(|| invalid("computed column input path is empty".to_string()))?;
+        let mut field = schema
+            .field(root)
+            .ok_or_else(|| invalid(format!("unknown computed column input '{field_path}'")))?;
+        let mut ancestors = Vec::new();
+        for name in rest {
+            ancestors.insert(0, field.id);
+            field = field
+                .children
+                .iter()
+                .find(|child| child.name == *name)
+                .ok_or_else(|| invalid(format!("unknown computed column input '{field_path}'")))?;
+        }
+        collect(field, parts, &ancestors, &mut out);
+    }
+    Ok(out)
+}
+
+/// Where one field's values come from in a fragment: the files and physical
+/// columns storing it, and the overlays overriding cells of it, newest last
+/// with the physical column and the cells each covers. A file stores the
+/// field under its own id or, packed, under an ancestor's; `ids` is the
+/// field's id followed by its ancestors'. Object identity is by base and
+/// path; field ids are left out, since a sibling column's rewrite re-labels
+/// them without touching a value.
+#[derive(Debug, PartialEq, Eq)]
+pub struct InputBasis {
+    files: Vec<(Option<u32>, String, i32)>,
+    overlays: Vec<(Option<u32>, String, i32, RoaringBitmap, u64)>,
+}
+
+pub fn input_basis(metadata: &Fragment, ids: &[i32]) -> Result<InputBasis> {
+    let column_of = |file: &DataFile| {
+        file.fields
+            .iter()
+            .position(|id| ids.contains(id))
+            .map(|pos| (pos, file.column_indices.get(pos).copied().unwrap_or(-1)))
+    };
+    let files = metadata
+        .files
+        .iter()
+        .filter_map(|file| {
+            column_of(file).map(|(_, column)| (file.base_id, file.path.clone(), column))
+        })
+        .collect();
+    let mut overlays = Vec::new();
+    for overlay in &metadata.overlays {
+        let Some((pos, column)) = column_of(&overlay.data_file) else {
+            continue;
+        };
+        overlays.push((
+            overlay.data_file.base_id,
+            overlay.data_file.path.clone(),
+            column,
+            overlay.coverage_for_field(pos)?.as_ref().clone(),
+            overlay.committed_version,
+        ));
+    }
+    Ok(InputBasis { files, overlays })
+}
+
+/// Identity of the input data a fragment currently holds: per input field,
+/// its storage basis. Deletions are left out: a deleted row is never
+/// computed, and the rows that stay keep their values. Physical identity,
+/// not content, so a rewrite that preserves values still reads as a change;
+/// compaction is followed separately.
+pub fn fragment_input_signature(fragment: &Fragment, inputs: &InputFields) -> Result<String> {
+    let mut parts = Vec::new();
+    for (path, ids) in inputs {
+        let basis = input_basis(fragment, ids)?;
+        parts.push(format!("{}={basis:?}", path.join(".")));
+    }
+    Ok(short_hash(&parts.join("|")))
+}
+
+fn signature_of(
+    dataset: &Dataset,
+    fragment_id: u32,
+    inputs: &InputFields,
+) -> Result<Option<String>> {
+    dataset
+        .get_fragment(fragment_id as usize)
+        .map(|fragment| fragment_input_signature(fragment.metadata(), inputs))
+        .transpose()
+}
+
+/// Signatures for `fragment_ids` as `dataset` currently holds them.
+pub fn signatures_for(
+    dataset: &Dataset,
+    fragment_ids: &[u32],
+    inputs: &InputFields,
+) -> Result<SignatureMap> {
+    let wanted: HashSet<u32> = fragment_ids.iter().copied().collect();
+    dataset
+        .get_fragments()
+        .iter()
+        .filter(|fragment| wanted.contains(&(fragment.id() as u32)))
+        .map(|fragment| {
+            Ok((
+                fragment.id() as u32,
+                fragment_input_signature(fragment.metadata(), inputs)?,
+            ))
+        })
+        .collect()
+}
+
+fn field_meta(dataset: &Dataset, column: &str, key: &str) -> Option<String> {
+    dataset
+        .schema()
+        .field(column)
+        .and_then(|field| field.metadata.get(key))
+        .cloned()
+}
+
+/// What the column's stored map says. The three cases are distinct and the
+/// callers act differently on each: see [`staleness_against`].
+pub enum StoredSignatures {
+    /// No map has ever been written for this column.
+    Absent,
+    Present(SignatureMap),
+    /// A map exists but cannot be read.
+    Unreadable,
+}
+
+/// Read the column's stored map. An unreadable map is a state, not an error:
+/// failing here would make the column permanently unrefreshable, and the
+/// unknown recomputes like every other unknown here.
+pub fn stored_signatures(dataset: &Dataset, column: &str) -> StoredSignatures {
+    let Some(encoded) = field_meta(dataset, column, SOURCE_SIGNATURE_META_KEY) else {
+        return StoredSignatures::Absent;
+    };
+    match serde_json::from_str(&encoded) {
+        Ok(map) => StoredSignatures::Present(map),
+        Err(error) => {
+            log::warn!(
+                "computed column '{column}' source signature map is unreadable ({error}); every fragment will be recomputed"
+            );
+            StoredSignatures::Unreadable
+        }
+    }
+}
+
+/// What a refresh must recompute beyond the null rows.
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+pub struct StalenessPlan {
+    /// The definition changed, so every row is stale whatever its signature.
+    pub recompute_all: bool,
+    /// Fragments whose inputs moved since they were computed, or that were
+    /// never recorded.
+    pub dirty: HashSet<u32>,
+    /// Fragments a compaction produced from fresh ones, at their current
+    /// signature: not dirty, and for the seal to record.
+    pub inherited: SignatureMap,
+}
+
+impl StalenessPlan {
+    pub fn is_dirty(&self, fragment_id: u32) -> bool {
+        self.recompute_all || self.dirty.contains(&fragment_id)
+    }
+}
+
+/// A version of the log since the stamp that bears on carrying freshness
+/// forward: an append, whose fragments hold no computed value yet, or a
+/// compaction's rewrite groups, consumed and produced fragment ids.
+enum Step {
+    Append,
+    Compaction(Vec<(Vec<u32>, Vec<u32>)>),
+}
+
+/// The step committed at `version`. A compaction's transaction records a
+/// new fragment before its id is assigned, so produced ids come from the
+/// version's manifest, matched by data file. Every other operation is
+/// skipped: a row-moving update rewrites the rows it moves, so it does not
+/// carry their inputs unchanged.
+async fn step_at(dataset: &Dataset, version: u64) -> Result<Option<Step>> {
+    let Some(transaction) = dataset.read_transaction_by_version(version).await? else {
+        return Ok(None);
+    };
+    let groups = match &transaction.operation {
+        Operation::Append { .. } => return Ok(Some(Step::Append)),
+        Operation::Rewrite { groups, .. } => groups,
+        _ => return Ok(None),
+    };
+    let at = dataset.checkout_version(version).await?;
+    let by_file: BTreeMap<(Option<u32>, String), u32> = at
+        .get_fragments()
+        .iter()
+        .flat_map(|fragment| {
+            let id = fragment.id() as u32;
+            fragment
+                .metadata()
+                .files
+                .iter()
+                .map(move |file| ((file.base_id, file.path.clone()), id))
+        })
+        .collect();
+    let compactions = groups
+        .iter()
+        .map(|group| {
+            let consumed = group.old_fragments.iter().map(|f| f.id as u32).collect();
+            let produced = group
+                .new_fragments
+                .iter()
+                .map(|fragment| {
+                    fragment
+                        .files
+                        .iter()
+                        .find_map(|file| by_file.get(&(file.base_id, file.path.clone())).copied())
+                        .ok_or_else(|| {
+                            invalid(format!(
+                                "a fragment added in version {version} is not in that version's manifest"
+                            ))
+                        })
+                })
+                .collect::<Result<Vec<u32>>>()?;
+            Ok((consumed, produced))
+        })
+        .collect::<Result<Vec<_>>>()?;
+    Ok(Some(Step::Compaction(compactions)))
+}
+
+/// Manifests since the stamp, each loaded once.
+struct Manifests<'a> {
+    dataset: &'a Dataset,
+    loaded: BTreeMap<u64, Dataset>,
+}
+
+impl Manifests<'_> {
+    async fn at(&mut self, version: u64) -> Result<&Dataset> {
+        if !self.loaded.contains_key(&version) {
+            let manifest = self.dataset.checkout_version(version).await?;
+            self.loaded.insert(version, manifest);
+        }
+        Ok(&self.loaded[&version])
+    }
+}
+
+/// Whether a fragment consumed by a compaction was created by an append
+/// since the stamp and its inputs never moved after: `current` is its
+/// signature just before the compaction. LanceDB's writes leave a computed
+/// column null (see `ensure_not_written`), but a raw append need not, so
+/// the rows it contributed to the product are checked to hold no value
+/// (`holds_values_in`) before the product inherits freshness.
+async fn appended_untouched(
+    manifests: &mut Manifests<'_>,
+    appends: &[u64],
+    fragment_id: u32,
+    current: Option<&String>,
+    inputs: &InputFields,
+) -> Result<bool> {
+    let Some(current) = current else {
+        return Ok(false);
+    };
+    for &version in appends.iter().rev() {
+        let Some(born) = signature_of(manifests.at(version).await?, fragment_id, inputs)? else {
+            continue;
+        };
+        if signature_of(manifests.at(version - 1).await?, fragment_id, inputs)?.is_some() {
+            // Live before this append: born earlier.
+            continue;
+        }
+        return Ok(&born == current);
+    }
+    Ok(false)
+}
+
+/// Whether `column` holds a value in any of `ranges`, offsets within the
+/// fragment. Compaction scans its sources in order, so the rows an appended
+/// source contributed sit at known offsets of the product.
+async fn holds_values_in(
+    dataset: &Dataset,
+    fragment_id: u32,
+    column: &str,
+    ranges: &[Range<u64>],
+) -> Result<bool> {
+    let Some(fragment) = dataset.get_fragment(fragment_id as usize) else {
+        return Ok(false);
+    };
+    let mut scanner = dataset.scan();
+    scanner
+        .with_fragments(vec![fragment.metadata().clone()])
+        .with_row_address()
+        .project(&[column])?
+        .filter(&format!(
+            "{} IS NOT NULL",
+            super::refresh::quote_identifier(column)
+        ))?;
+    let mut batches = scanner.try_into_stream().await?;
+    while let Some(batch) = batches.try_next().await? {
+        let addresses = batch
+            .column_by_name(ROW_ADDR)
+            .and_then(|column| column.as_any().downcast_ref::<UInt64Array>())
+            .ok_or_else(|| invalid("row addresses missing from a freshness scan".to_string()))?;
+        if addresses
+            .iter()
+            .flatten()
+            .map(|address| address & 0xFFFF_FFFF)
+            .any(|offset| ranges.iter().any(|range| range.contains(&offset)))
+        {
+            return Ok(true);
+        }
+    }
+    Ok(false)
+}
+
+/// Compaction copies inputs verbatim, so a fragment it produced from
+/// recorded fragments whose inputs had not moved is as fresh as they were,
+/// and a fragment an append created since the stamp, untouched after and
+/// unfilled in the product, changes nothing (`appended_untouched`).
+/// Followed from the version the map was recorded at through every
+/// compaction since, so a chain of them carries too. Returns the produced
+/// fragments' signatures at production; the caller compares each with the
+/// current manifest, which catches anything written to them afterwards.
+async fn carried_forward(
+    dataset: &Dataset,
+    column: &str,
+    stored: &SignatureMap,
+    inputs: &InputFields,
+) -> Result<SignatureMap> {
+    let Some(recorded_at) = field_meta(dataset, column, RECORDED_AT_VERSION_META_KEY)
+        .and_then(|version| version.parse::<u64>().ok())
+    else {
+        return Ok(SignatureMap::new());
+    };
+    let to = dataset.version().version;
+    if to <= recorded_at || to - recorded_at > MAX_CARRY_FORWARD_VERSIONS {
+        return Ok(SignatureMap::new());
+    }
+    let mut fresh = stored.clone();
+    let mut inherited = SignatureMap::new();
+    let mut appends = Vec::new();
+    let mut manifests = Manifests {
+        dataset,
+        loaded: BTreeMap::new(),
+    };
+    for version in (recorded_at + 1)..=to {
+        let compactions = match step_at(dataset, version).await? {
+            None => continue,
+            Some(Step::Append) => {
+                appends.push(version);
+                continue;
+            }
+            Some(Step::Compaction(compactions)) => compactions,
+        };
+        for (consumed, produced) in compactions {
+            // Each source's signature and live rows just before the
+            // compaction; live rows place its contribution in the product.
+            let mut sources = Vec::with_capacity(consumed.len());
+            {
+                let before = manifests.at(version - 1).await?;
+                for id in &consumed {
+                    let live = match before.get_fragment(*id as usize) {
+                        Some(fragment) => fragment.count_rows(None).await? as u64,
+                        None => 0,
+                    };
+                    sources.push((*id, signature_of(before, *id, inputs)?, live));
+                }
+            }
+            let mut all_fresh = !consumed.is_empty();
+            let mut appended = Vec::new();
+            let mut offset = 0u64;
+            for (id, current, live) in sources {
+                let recorded = matches!(
+                    (fresh.get(&id), current.as_ref()),
+                    (Some(recorded), Some(current)) if recorded == current
+                );
+                if !recorded {
+                    if appended_untouched(&mut manifests, &appends, id, current.as_ref(), inputs)
+                        .await?
+                    {
+                        appended.push(offset..offset + live);
+                    } else {
+                        all_fresh = false;
+                        break;
+                    }
+                }
+                offset += live;
+            }
+            if !all_fresh {
+                continue;
+            }
+            let after = manifests.at(version).await?;
+            // The products in order, each with its share of the appended
+            // rows; a value there was supplied by the append, not computed.
+            let mut base = 0u64;
+            let mut unfilled = true;
+            for id in &produced {
+                let rows = match after.get_fragment(*id as usize) {
+                    Some(fragment) => fragment.count_rows(None).await? as u64,
+                    None => 0,
+                };
+                let local: Vec<Range<u64>> = appended
+                    .iter()
+                    .filter(|range| range.start < base + rows && range.end > base)
+                    .map(|range| range.start.max(base) - base..range.end.min(base + rows) - base)
+                    .collect();
+                if !local.is_empty() && holds_values_in(after, *id, column, &local).await? {
+                    unfilled = false;
+                    break;
+                }
+                base += rows;
+            }
+            if !unfilled {
+                continue;
+            }
+            for id in &produced {
+                if let Some(signature) = signature_of(after, *id, inputs)? {
+                    fresh.insert(*id, signature.clone());
+                    inherited.insert(*id, signature);
+                }
+            }
+        }
+    }
+    Ok(inherited)
+}
+
+/// Decide what is stale, from the manifest alone.
+///
+/// A column with no map at all was declared before signatures existed. It
+/// keeps its null-fill behavior until its first stamp enrolls it (see
+/// [`record_freshness`]). A map that omits a fragment is authoritative: the
+/// fragment's input state is unknown, and unknown recomputes -- unless a
+/// compaction of fresh fragments produced it, which is followed. Lineage is
+/// read only when a live fragment has no entry, so a plan over a recorded
+/// table costs no transaction reads.
+pub async fn staleness_against(
+    dataset: &Dataset,
+    column: &str,
+    definition_version: &str,
+    inputs: &InputFields,
+) -> Result<StalenessPlan> {
+    let stored = match stored_signatures(dataset, column) {
+        StoredSignatures::Absent => return Ok(StalenessPlan::default()),
+        StoredSignatures::Unreadable => {
+            return Ok(StalenessPlan {
+                recompute_all: true,
+                ..Default::default()
+            });
+        }
+        StoredSignatures::Present(stored) => stored,
+    };
+    let stored_version = field_meta(dataset, column, DEFINITION_VERSION_META_KEY);
+    if stored_version.is_some_and(|version| version != definition_version) {
+        return Ok(StalenessPlan {
+            recompute_all: true,
+            ..Default::default()
+        });
+    }
+    let unrecorded = dataset
+        .get_fragments()
+        .iter()
+        .any(|fragment| !stored.contains_key(&(fragment.id() as u32)));
+    let mut inherited = if unrecorded {
+        carried_forward(dataset, column, &stored, inputs).await?
+    } else {
+        SignatureMap::new()
+    };
+    let mut dirty = HashSet::new();
+    let mut live = HashSet::new();
+    for fragment in dataset.get_fragments() {
+        let id = fragment.id() as u32;
+        live.insert(id);
+        let current = fragment_input_signature(fragment.metadata(), inputs)?;
+        if stored.get(&id).or_else(|| inherited.get(&id)) != Some(&current) {
+            dirty.insert(id);
+        }
+    }
+    inherited.retain(|id, _| live.contains(id) && !dirty.contains(id));
+    Ok(StalenessPlan {
+        recompute_all: false,
+        dirty,
+        inherited,
+    })
+}
+
+/// What [`record_freshness`] wrote: the table version the stamp landed at,
+/// if it wrote one, and the entries recorded and dropped for having moved.
+#[derive(Debug, Default, PartialEq, Eq)]
+pub struct FreshnessRecord {
+    pub version: Option<u64>,
+    pub recorded: usize,
+    pub moved: usize,
+}
+
+/// Record, on `column`, the input state its fragments were computed from.
+///
+/// `computed` is what this refresh computed in full, signed at the version
+/// the values were read from. A column with no map yet was declared before
+/// signatures existed; `pinned`, the version the refresh planned against,
+/// is then the baseline: every fragment live there is trusted as it stood,
+/// the null-fill contract its values were written under. Otherwise the
+/// staleness decided on `pinned` supplies what compactions since the last
+/// stamp carried forward. Either
+/// way an entry is recorded only if `latest` still holds that input state --
+/// an input write can rebase under the output commit -- so a fragment whose
+/// inputs moved stays unrecorded and is recomputed by the next refresh.
+///
+/// Written once per refresh, after its data commit.
+pub async fn record_freshness(
+    latest: &mut Dataset,
+    pinned: Option<(&Dataset, &StalenessPlan)>,
+    column: &str,
+    definition_version: &str,
+    inputs: &InputFields,
+    computed: SignatureMap,
+) -> Result<FreshnessRecord> {
+    let absent = matches!(stored_signatures(latest, column), StoredSignatures::Absent);
+    let mut entries = SignatureMap::new();
+    if let Some((pinned, staleness)) = pinned {
+        if absent {
+            let all: Vec<u32> = pinned
+                .get_fragments()
+                .iter()
+                .map(|fragment| fragment.id() as u32)
+                .collect();
+            entries = signatures_for(pinned, &all, inputs)?;
+        } else {
+            entries = staleness.inherited.clone();
+        }
+    }
+    entries.extend(computed);
+    if entries.is_empty() && !absent {
+        return Ok(FreshnessRecord::default());
+    }
+    let fragments: Vec<u32> = entries.keys().copied().collect();
+    let current = signatures_for(latest, &fragments, inputs)?;
+    let verified: SignatureMap = entries
+        .into_iter()
+        .filter(|(fragment_id, signature)| current.get(fragment_id) == Some(signature))
+        .collect();
+    let recorded = verified.len();
+    let version = write_signatures(latest, column, definition_version, verified).await?;
+    Ok(FreshnessRecord {
+        version: Some(version),
+        recorded,
+        moved: fragments.len() - recorded,
+    })
+}
+
+/// Merge `entries` into the column's stored map and stamp the definition and
+/// the version the map now describes. Merges rather than replaces: the
+/// entries cover only the fragments this refresh wrote, and every fragment it
+/// skipped keeps the entry an earlier one left. Entries for fragments no
+/// longer in the manifest are dropped, so compaction cannot grow the map
+/// without bound. Returns the version the stamp landed at.
+pub async fn write_signatures(
+    dataset: &mut Dataset,
+    column: &str,
+    definition_version: &str,
+    entries: SignatureMap,
+) -> Result<u64> {
+    let live: HashSet<u32> = dataset
+        .get_fragments()
+        .iter()
+        .map(|fragment| fragment.id() as u32)
+        .collect();
+    // An unreadable map is discarded rather than merged: nothing in it can be
+    // trusted, and its fragments recompute until a later refresh records them.
+    let mut merged = match stored_signatures(dataset, column) {
+        StoredSignatures::Present(stored) => stored,
+        StoredSignatures::Absent | StoredSignatures::Unreadable => SignatureMap::new(),
+    };
+    merged.extend(entries);
+    merged.retain(|fragment_id, _| live.contains(fragment_id));
+    let encoded = serde_json::to_string(&merged).map_err(|e| Error::Runtime {
+        message: format!("encoding {column} signatures: {e}"),
+    })?;
+    let describes = dataset.version().version;
+    dataset
+        .update_field_metadata()
+        .update(
+            column,
+            [
+                (SOURCE_SIGNATURE_META_KEY.to_string(), encoded),
+                (
+                    DEFINITION_VERSION_META_KEY.to_string(),
+                    definition_version.to_string(),
+                ),
+                (
+                    RECORDED_AT_VERSION_META_KEY.to_string(),
+                    describes.to_string(),
+                ),
+            ],
+        )?
+        .await?;
+    Ok(dataset.version().version)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::sync::Arc;
+
+    use arrow_array::RecordBatchIterator;
+    use arrow_schema::{DataType, Field as ArrowField, Schema as ArrowSchema};
+    use lance::dataset::{
+        MergeInsertBuilder, MergeInsertWriteMode, NewColumnTransform, WhenMatched, WhenNotMatched,
+        WriteMode, WriteParams,
+    };
+    use lance_file::version::ConcreteFileVersion;
+    use lance_table::format::overlay::{DataOverlayFile, OverlayCoverage};
+
+    const COLUMN: &str = "doubled";
+    const EXPRESSION: &str = "value * 2";
+
+    /// Two fragments of 50 rows, `id` and `value`, with `doubled` declared
+    /// all-null against `value`, tracked from birth.
+    async fn table(uri: &str) -> Dataset {
+        let batch = arrow_array::record_batch!(
+            ("id", Int32, (0..100).collect::<Vec<i32>>()),
+            ("value", Int32, (0..100).collect::<Vec<i32>>())
+        )
+        .unwrap();
+        let schema = batch.schema();
+        let mut dataset = Dataset::write(
+            RecordBatchIterator::new(vec![Ok(batch)], schema),
+            uri,
+            Some(WriteParams {
+                mode: WriteMode::Create,
+                max_rows_per_file: 50,
+                ..Default::default()
+            }),
+        )
+        .await
+        .unwrap();
+        let mut metadata = std::collections::HashMap::from([
+            (
+                crate::table::computed_columns::COMPUTED_COLUMN_META_KEY.to_string(),
+                "true".to_string(),
+            ),
+            (
+                crate::table::computed_columns::EXPRESSION_META_KEY.to_string(),
+                EXPRESSION.to_string(),
+            ),
+        ]);
+        metadata.insert(SOURCE_SIGNATURE_META_KEY.to_string(), "{}".to_string());
+        dataset
+            .add_columns(
+                NewColumnTransform::AllNulls(Arc::new(ArrowSchema::new(vec![
+                    ArrowField::new(COLUMN, DataType::Int64, true).with_metadata(metadata),
+                ]))),
+                None,
+                None,
+            )
+            .await
+            .unwrap();
+        dataset
+    }
+
+    fn inputs(dataset: &Dataset) -> InputFields {
+        fields_for_paths(dataset.schema(), &["value".to_string()]).unwrap()
+    }
+
+    async fn plan(dataset: &Dataset) -> StalenessPlan {
+        staleness_against(
+            dataset,
+            COLUMN,
+            &definition_version(EXPRESSION),
+            &inputs(dataset),
+        )
+        .await
+        .unwrap()
+    }
+
+    async fn stamp_all(dataset: &mut Dataset) {
+        let ids = inputs(dataset);
+        let frags: Vec<u32> = dataset
+            .get_fragments()
+            .iter()
+            .map(|f| f.id() as u32)
+            .collect();
+        let entries = signatures_for(dataset, &frags, &ids).unwrap();
+        write_signatures(dataset, COLUMN, &definition_version(EXPRESSION), entries)
+            .await
+            .unwrap();
+    }
+
+    /// Strip the refresh's own keys: a column from before signatures existed.
+    async fn make_legacy(dataset: &mut Dataset) {
+        let declaration = dataset
+            .schema()
+            .field(COLUMN)
+            .unwrap()
+            .metadata
+            .iter()
+            .filter(|(key, _)| !key.starts_with("computed_refresh."))
+            .map(|(key, value)| (key.clone(), value.clone()))
+            .collect::<Vec<_>>();
+        dataset
+            .update_field_metadata()
+            .replace(COLUMN, declaration)
+            .unwrap()
+            .await
+            .unwrap();
+        assert!(matches!(
+            stored_signatures(dataset, COLUMN),
+            StoredSignatures::Absent
+        ));
+    }
+
+    /// Rewrite `value` of the row with `id` in place: a partial merge-insert
+    /// attaches a new column file to the row's fragment, keeping its id.
+    async fn rewrite_value(dataset: &mut Dataset, id: i32) {
+        let batch =
+            arrow_array::record_batch!(("id", Int32, [id]), ("value", Int32, [1000])).unwrap();
+        let schema = batch.schema();
+        let mut builder =
+            MergeInsertBuilder::try_new(Arc::new(dataset.clone()), vec!["id".to_string()]).unwrap();
+        builder
+            .when_matched(WhenMatched::UpdateAll)
+            .when_not_matched(WhenNotMatched::DoNothing)
+            .write_mode(MergeInsertWriteMode::RewriteColumns);
+        let (updated, _) = builder
+            .try_build()
+            .unwrap()
+            .execute_reader(RecordBatchIterator::new([Ok(batch)], schema))
+            .await
+            .unwrap();
+        *dataset = (*updated).clone();
+    }
+
+    async fn compact(dataset: &mut Dataset) -> u32 {
+        lance::dataset::optimize::compact_files(
+            dataset,
+            lance::dataset::optimize::CompactionOptions {
+                target_rows_per_fragment: 1000,
+                ..Default::default()
+            },
+            None,
+        )
+        .await
+        .unwrap();
+        dataset.checkout_latest().await.unwrap();
+        dataset.get_fragments()[0].id() as u32
+    }
+
+    /// `a.b` under `root` and a nested `a` -> `b` are different columns with
+    /// two bases; a dotted-string key would fold them into one.
+    #[test]
+    fn a_dotted_field_name_is_not_a_nested_path() {
+        let leaf = |name: &str| ArrowField::new(name, DataType::Int32, true);
+        let root = ArrowField::new(
+            "root",
+            DataType::Struct(
+                vec![
+                    ArrowField::new("a", DataType::Struct(vec![leaf("b")].into()), true),
+                    leaf("a.b"),
+                ]
+                .into(),
+            ),
+            true,
+        );
+        let schema = LanceSchema::try_from(&ArrowSchema::new(vec![root])).unwrap();
+        let ids = fields_for_paths(&schema, &["root".to_string()]).unwrap();
+        let path = |parts: &[&str]| parts.iter().map(|p| p.to_string()).collect::<Vec<_>>();
+        let nested = &ids[&path(&["root", "a", "b"])];
+        let dotted = &ids[&path(&["root", "a.b"])];
+        assert_ne!(nested[0], dotted[0], "{ids:?}");
+        assert_eq!(ids.len(), 4, "{ids:?}");
+    }
+
+    /// A packed file records the physical column under an ancestor's id, so a
+    /// nested input's basis is found through its ancestors.
+    #[test]
+    fn a_packed_nested_input_has_a_file_basis() {
+        let word_count = ArrowField::new("word_count", DataType::Int32, true);
+        let metrics = ArrowField::new("metrics", DataType::Struct(vec![word_count].into()), true);
+        let analysis = ArrowField::new("analysis", DataType::Struct(vec![metrics].into()), true);
+        let schema = LanceSchema::try_from(&ArrowSchema::new(vec![analysis])).unwrap();
+        let ids = fields_for_paths(&schema, &["analysis.metrics.word_count".to_string()]).unwrap();
+        let word_count = &ids[&["analysis", "metrics", "word_count"]
+            .map(String::from)
+            .to_vec()];
+        let analysis = schema.field("analysis").unwrap().id;
+        assert_eq!(word_count.last(), Some(&analysis), "{word_count:?}");
+        let mut fragment = Fragment::new(0);
+        fragment.files.push(DataFile::new(
+            "packed.lance",
+            vec![analysis],
+            vec![3],
+            ConcreteFileVersion::V2_2,
+            None,
+            None,
+        ));
+        let basis = input_basis(&fragment, word_count).unwrap();
+        assert_eq!(basis.files, vec![(None, "packed.lance".to_string(), 3)]);
+    }
+
+    /// An overlay that stores the input in another physical column of the
+    /// same object is a different basis.
+    #[test]
+    fn an_overlay_column_remap_changes_the_basis() {
+        let overlay = |column: i32| {
+            let mut fragment = Fragment::new(0);
+            fragment.overlays.push(DataOverlayFile {
+                data_file: DataFile::new(
+                    "overlay.lance",
+                    vec![7],
+                    vec![column],
+                    ConcreteFileVersion::V2_2,
+                    None,
+                    None,
+                ),
+                coverage: OverlayCoverage::dense(RoaringBitmap::from_iter([0u32])),
+                committed_version: 2,
+            });
+            input_basis(&fragment, &[7]).unwrap()
+        };
+        assert_ne!(overlay(0), overlay(1));
+        assert_eq!(overlay(0), overlay(0));
+    }
+
+    /// A freshly declared column is tracked from birth: every fragment is
+    /// unrecorded, so every fragment is stale until a refresh records it.
+    #[tokio::test]
+    async fn a_declared_column_is_stale_until_recorded() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        assert_eq!(plan(&dataset).await.dirty, HashSet::from([0, 1]));
+        stamp_all(&mut dataset).await;
+        assert_eq!(plan(&dataset).await, StalenessPlan::default());
+    }
+
+    /// The signature answers "did my inputs move": a write to any other
+    /// column, the computed column included, leaves it alone.
+    #[tokio::test]
+    async fn an_unrelated_column_rewrite_leaves_the_signature_alone() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        let ids = inputs(&dataset);
+        let before = signatures_for(&dataset, &[0, 1], &ids).unwrap();
+        dataset
+            .add_columns(
+                NewColumnTransform::SqlExpressions(vec![("extra".into(), "id * 3".into())]),
+                None,
+                None,
+            )
+            .await
+            .unwrap();
+        assert_eq!(before, signatures_for(&dataset, &[0, 1], &ids).unwrap());
+    }
+
+    /// An in-place write to one fragment's input keeps every fragment id, so
+    /// only the signature can notice -- and on that fragment alone.
+    #[tokio::test]
+    async fn an_in_place_input_change_dirties_only_its_own_fragment() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        rewrite_value(&mut dataset, 60).await;
+        let stale = plan(&dataset).await;
+        assert_eq!(stale.dirty, HashSet::from([1]), "{stale:?}");
+    }
+
+    /// A deleted row is never computed and the rows that stay keep their
+    /// values, so a delete dirties nothing.
+    #[tokio::test]
+    async fn a_delete_dirties_nothing() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        dataset.delete("id >= 60 AND id < 70").await.unwrap();
+        assert_eq!(plan(&dataset).await, StalenessPlan::default());
+    }
+
+    /// A definition change makes every row stale whatever the signatures say.
+    #[tokio::test]
+    async fn a_definition_change_recomputes_every_row() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        let rebound = staleness_against(&dataset, COLUMN, "other", &inputs(&dataset))
+            .await
+            .unwrap();
+        assert!(rebound.recompute_all);
+    }
+
+    /// A column declared before signatures existed carries no map, and keeps
+    /// null-fill behavior until its first stamp enrolls it -- at the pinned
+    /// state, not the latest: an input that moved in between is left out.
+    #[tokio::test]
+    async fn a_first_stamp_enrolls_an_older_column_as_it_stood() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        make_legacy(&mut dataset).await;
+        assert_eq!(plan(&dataset).await, StalenessPlan::default());
+        let ids = inputs(&dataset);
+        let pinned = dataset.clone();
+        let staleness = plan(&pinned).await;
+        rewrite_value(&mut dataset, 60).await;
+        let record = record_freshness(
+            &mut dataset,
+            Some((&pinned, &staleness)),
+            COLUMN,
+            &definition_version(EXPRESSION),
+            &ids,
+            SignatureMap::new(),
+        )
+        .await
+        .unwrap();
+        assert_eq!((record.recorded, record.moved), (1, 1));
+        assert_eq!(record.version, Some(dataset.version().version));
+        assert_eq!(plan(&dataset).await.dirty, HashSet::from([1]));
+    }
+
+    /// Append `count` rows after `first_id`, the computed column null as a
+    /// write must leave it; returns the new fragment's id.
+    async fn append_rows(dataset: &mut Dataset, first_id: i32, count: i32) -> u32 {
+        let batch = arrow_array::record_batch!(
+            (
+                "id",
+                Int32,
+                (first_id..first_id + count).collect::<Vec<i32>>()
+            ),
+            (
+                "value",
+                Int32,
+                (first_id..first_id + count).collect::<Vec<i32>>()
+            ),
+            ("doubled", Int64, vec![None::<i64>; count as usize])
+        )
+        .unwrap();
+        let schema = batch.schema();
+        *dataset = Dataset::write(
+            RecordBatchIterator::new(vec![Ok(batch)], schema),
+            dataset.uri(),
+            Some(WriteParams {
+                mode: WriteMode::Append,
+                ..Default::default()
+            }),
+        )
+        .await
+        .unwrap();
+        dataset.get_fragments().last().unwrap().id() as u32
+    }
+
+    /// A fragment an append created since the stamp holds no computed value,
+    /// so a compaction folding it into recorded fragments produces a fresh
+    /// fragment: nothing recomputes, and the null fill covers the new rows.
+    #[tokio::test]
+    async fn a_compaction_folding_an_untouched_appended_fragment_carries_freshness_forward() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        let appended = append_rows(&mut dataset, 100, 10).await;
+        assert_eq!(plan(&dataset).await.dirty, HashSet::from([appended]));
+        let compacted = compact(&mut dataset).await;
+        let stale = plan(&dataset).await;
+        assert!(stale.dirty.is_empty(), "{stale:?}");
+        assert_eq!(
+            stale.inherited.keys().copied().collect::<Vec<u32>>(),
+            vec![compacted]
+        );
+    }
+
+    /// The same fragment with an input rewritten after the append is not
+    /// neutral: what it holds may have been computed from the older input.
+    #[tokio::test]
+    async fn a_compaction_folding_an_appended_fragment_whose_input_moved_recomputes() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        append_rows(&mut dataset, 100, 10).await;
+        rewrite_value(&mut dataset, 105).await;
+        let compacted = compact(&mut dataset).await;
+        let stale = plan(&dataset).await;
+        assert_eq!(stale.dirty, HashSet::from([compacted]), "{stale:?}");
+        assert!(stale.inherited.is_empty());
+    }
+
+    /// A raw append can supply a computed value LanceDB's own writes never
+    /// do. The product holds it where the appended rows landed, so the
+    /// compaction is not carried: the value is recomputed, not certified.
+    #[tokio::test]
+    async fn a_compaction_folding_an_appended_fragment_with_values_recomputes() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        let batch = arrow_array::record_batch!(
+            ("id", Int32, [100, 101]),
+            ("value", Int32, [100, 101]),
+            ("doubled", Int64, [None, Some(999_i64)])
+        )
+        .unwrap();
+        let schema = batch.schema();
+        dataset = Dataset::write(
+            RecordBatchIterator::new(vec![Ok(batch)], schema),
+            dataset.uri(),
+            Some(WriteParams {
+                mode: WriteMode::Append,
+                ..Default::default()
+            }),
+        )
+        .await
+        .unwrap();
+        let compacted = compact(&mut dataset).await;
+        let stale = plan(&dataset).await;
+        assert_eq!(stale.dirty, HashSet::from([compacted]), "{stale:?}");
+        assert!(stale.inherited.is_empty());
+    }
+
+    /// Compaction copies inputs unchanged: a fragment it produced from
+    /// recorded, unmoved fragments is fresh, and the seal records it. One
+    /// produced from a fragment whose inputs had moved is not.
+    #[tokio::test]
+    async fn a_compaction_of_fresh_fragments_carries_freshness_forward() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        let compacted = compact(&mut dataset).await;
+        let stale = plan(&dataset).await;
+        assert!(stale.dirty.is_empty(), "{stale:?}");
+        assert_eq!(
+            stale.inherited.keys().copied().collect::<Vec<u32>>(),
+            vec![compacted]
+        );
+        let ids = inputs(&dataset);
+        let pinned = dataset.clone();
+        let record = record_freshness(
+            &mut dataset,
+            Some((&pinned, &stale)),
+            COLUMN,
+            &definition_version(EXPRESSION),
+            &ids,
+            SignatureMap::new(),
+        )
+        .await
+        .unwrap();
+        assert_eq!(record.recorded, 1);
+        let StoredSignatures::Present(stored) = stored_signatures(&dataset, COLUMN) else {
+            panic!("stamped");
+        };
+        assert_eq!(
+            stored.keys().copied().collect::<Vec<u32>>(),
+            vec![compacted]
+        );
+
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        stamp_all(&mut dataset).await;
+        rewrite_value(&mut dataset, 60).await;
+        let compacted = compact(&mut dataset).await;
+        let stale = plan(&dataset).await;
+        assert_eq!(stale.dirty, HashSet::from([compacted]), "{stale:?}");
+        assert!(stale.inherited.is_empty());
+    }
+
+    /// A recorded fragment whose signature no longer matches, or a fragment
+    /// the map omits without a compaction to explain it, is stale.
+    #[tokio::test]
+    async fn a_fragment_missing_from_the_stored_map_is_stale() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        let ids = inputs(&dataset);
+        let partial = signatures_for(&dataset, &[0], &ids).unwrap();
+        write_signatures(
+            &mut dataset,
+            COLUMN,
+            &definition_version(EXPRESSION),
+            partial,
+        )
+        .await
+        .unwrap();
+        assert_eq!(plan(&dataset).await.dirty, HashSet::from([1]));
+    }
+
+    /// An unreadable map recomputes everything rather than failing the
+    /// refresh, and the next stamp replaces it.
+    #[tokio::test]
+    async fn an_unreadable_stored_map_recomputes_rather_than_failing() {
+        let dir = tempfile::tempdir().unwrap();
+        let mut dataset = table(dir.path().to_str().unwrap()).await;
+        dataset
+            .update_field_metadata()
+            .update(
+                COLUMN,
+                [(SOURCE_SIGNATURE_META_KEY.to_string(), "{".to_string())],
+            )
+            .unwrap()
+            .await
+            .unwrap();
+        assert!(plan(&dataset).await.recompute_all);
+        stamp_all(&mut dataset).await;
+        assert_eq!(plan(&dataset).await, StalenessPlan::default());
+    }
+}
diff --git a/rust/lancedb/src/table/refresh.rs b/rust/lancedb/src/table/refresh.rs
index 511fce8ff..64d519a58 100644
--- a/rust/lancedb/src/table/refresh.rs
+++ b/rust/lancedb/src/table/refresh.rs
@@ -3,9 +3,9 @@
 
 //! Filling computed columns.
 //!
-//! A row without a value gets one; a row that has one keeps it. Refresh is
-//! therefore idempotent and does not observe input mutation -- once a row is
-//! filled, changing what the expression reads leaves the stored result alone.
+//! A row without a value gets one; a row that has one keeps it unless its
+//! fragment's inputs moved since it was computed, which `freshness` decides
+//! from the manifest and stamps after every fill.
 //!
 //! A column's computed inputs are filled first -- the dependency graph is
 //! walked once, each reachable column filled once in dependency order, each
@@ -31,6 +31,7 @@
 
 use std::collections::HashSet;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
 
 use arrow_array::{
     Array, ArrayRef, BooleanArray, LargeBinaryArray, RecordBatch, RecordBatchOptions, StructArray,
@@ -48,6 +49,7 @@ use lance_core::datatypes::{BlobHandling, Schema as LanceSchema};
 use serde::{Deserialize, Serialize};
 
 use super::computed_columns::{BoundExpression, ComputedColumnKind, computed_column_from_field};
+use super::freshness::{self, SignatureMap, StalenessPlan};
 use super::{BaseTable, NativeTable};
 use crate::job::Job;
 use crate::{Error, Result};
@@ -110,28 +112,81 @@ async fn execute_refresh_column_with_source(
     };
     let output_is_blob = field.is_blob_v2();
 
+    // Which fragments the null filter cannot speak for: their inputs moved
+    // since they were computed, or the definition did. Decided once, from
+    // the manifest the values are read from.
+    let inputs = freshness::fields_for_paths(dataset.schema(), &bound.inputs)?;
+    let definition = freshness::definition_version(&expression);
+    let staleness = freshness::staleness_against(&dataset, column, &definition, &inputs).await?;
+
     let mut rows_filled = 0u64;
     let mut replacements = Vec::new();
+    // Fragments this refresh computed in full, signed at the version read.
+    let mut computed = SignatureMap::new();
     for fragment in dataset.get_fragments() {
-        let gained = count_fragment_gains(&dataset, &fragment, &bound, column).await?;
-        if gained == 0 {
-            continue;
+        let fragment_id = u32::try_from(fragment.id()).map_err(|_| Error::Runtime {
+            message: format!("fragment id {} does not fit a signature map", fragment.id()),
+        })?;
+        // A recompute rewrites every live row, so it is staged without the
+        // probe and counted as it fills; a null fill probes first, since a
+        // fragment with nothing to gain is not worth a write.
+        let recompute = staleness.is_dirty(fragment_id);
+        let whole = recompute || {
+            let (gained, unfilled) =
+                count_fragment_gains(&dataset, &fragment, &bound, column).await?;
+            if gained == 0 {
+                continue;
+            }
+            rows_filled += gained;
+            unfilled == u64::try_from(fragment.count_rows(None).await?).unwrap_or(u64::MAX)
+        };
+        if whole {
+            computed.insert(
+                fragment_id,
+                freshness::fragment_input_signature(fragment.metadata(), &inputs)?,
+            );
         }
-        rows_filled += gained;
-        let values =
-            fill_stream(&dataset, &fragment, bound.clone(), column, output_is_blob).await?;
+        let gained = Arc::new(AtomicU64::new(0));
+        let values = fill_stream(
+            &dataset,
+            &fragment,
+            bound.clone(),
+            column,
+            output_is_blob,
+            recompute,
+            gained.clone(),
+        )
+        .await?;
         replacements.push(fragment.write_columns(values, &column_schema).await?);
+        if recompute {
+            rows_filled += gained.load(Ordering::Relaxed);
+        }
     }
 
     let source_version = dataset.version().version;
     if replacements.is_empty() {
+        // Nothing to fill; the stamp may still have something to record -- a
+        // column not yet enrolled, or fragments a compaction carried.
+        let mut latest = (*dataset).clone();
+        let stamped = record(
+            &mut latest,
+            (&dataset, &staleness),
+            column,
+            &definition,
+            &inputs,
+            computed,
+        )
+        .await;
+        if stamped.is_some() {
+            table.dataset.update(latest);
+        }
         return Ok(RefreshExecution {
             result: RefreshColumnResult {
                 rows_filled: 0,
-                version: source_version,
+                version: stamped.unwrap_or(source_version),
             },
             source_version,
-            published_version: None,
+            published_version: stamped,
         });
     }
 
@@ -149,7 +204,17 @@ async fn execute_refresh_column_with_source(
     )
     .await?;
 
-    let version = new_dataset.version().version;
+    let mut new_dataset = new_dataset;
+    let version = record(
+        &mut new_dataset,
+        (&dataset, &staleness),
+        column,
+        &definition,
+        &inputs,
+        computed,
+    )
+    .await
+    .unwrap_or(new_dataset.version().version);
     table.dataset.update(new_dataset);
     Ok(RefreshExecution {
         result: RefreshColumnResult {
@@ -161,6 +226,31 @@ async fn execute_refresh_column_with_source(
     })
 }
 
+/// Stamp the input state the refresh computed from (see
+/// [`freshness::record_freshness`]); the version the stamp landed at, which
+/// is the last one the refresh wrote. Never fails the refresh: the values
+/// are committed, and a missing stamp only costs a recompute next time.
+async fn record(
+    latest: &mut Dataset,
+    pinned: (&Dataset, &StalenessPlan),
+    column: &str,
+    definition: &str,
+    inputs: &freshness::InputFields,
+    computed: SignatureMap,
+) -> Option<u64> {
+    match freshness::record_freshness(latest, Some(pinned), column, definition, inputs, computed)
+        .await
+    {
+        Ok(record) => record.version,
+        Err(error) => {
+            log::warn!(
+                "could not record the input state computed column '{column}' was refreshed from ({error}); its fragments will recompute on the next refresh"
+            );
+            None
+        }
+    }
+}
+
 /// Refuse while a computed input still has rows a refresh of it would fill:
 /// read now, its placeholder null would be evaluated as a value and kept.
 async fn ensure_inputs_filled(
@@ -188,7 +278,9 @@ async fn ensure_inputs_filled(
         let input_bound = super::computed_columns::bind(schema.clone(), input, expression)?;
         let mut unfilled = 0u64;
         for fragment in dataset.get_fragments() {
-            unfilled += count_fragment_gains(dataset, &fragment, &input_bound, input).await?;
+            unfilled += count_fragment_gains(dataset, &fragment, &input_bound, input)
+                .await?
+                .0;
         }
         if unfilled > 0 {
             return Err(Error::InvalidInput {
@@ -436,32 +528,35 @@ fn blob_array_from_binary(
 /// Scans only the unfilled live rows -- deleted rows never reach the
 /// expression here, the filter having already excluded them -- and counts the
 /// non-null results. Exact, so it is both the staging decision and the
-/// fragment's contribution to `rows_filled`.
+/// fragment's contribution to `rows_filled`. Returns the gains and the rows
+/// scanned.
 async fn count_fragment_gains(
     dataset: &Dataset,
     fragment: &FileFragment,
     bound: &BoundExpression,
     column: &str,
-) -> Result<u64> {
+) -> Result<(u64, u64)> {
     let mut scanner = dataset.scan();
     scanner
         .with_fragments(vec![fragment.metadata().clone()])
         .with_row_id()
-        .filter(&format!("{} IS NULL", quote_identifier(column)))?
-        .project(&bound.roots)?;
+        .project(&bound.roots)?
+        .filter(&format!("{} IS NULL", quote_identifier(column)))?;
     configure_blob_inputs(&mut scanner, dataset.schema(), bound, None)?;
 
     let mut gained = 0u64;
+    let mut considered = 0u64;
     let mut batches = scanner.try_into_stream().await?;
     while let Some(batch) = batches.try_next().await? {
         let evaluated = evaluate(bound, &evaluation_batch(&batch, bound, None)?)?;
         gained += (batch.num_rows() - evaluated.null_count()) as u64;
+        considered += batch.num_rows() as u64;
     }
-    Ok(gained)
+    Ok((gained, considered))
 }
 
 /// Stream one fragment's column in physical order, filling the unfilled live
-/// rows and keeping every other value.
+/// rows -- every live row, for a recompute -- and keeping every other value.
 ///
 /// Deleted rows are carried through so the values line up positionally with
 /// the fragment's data files; they are never read back, but the column file
@@ -472,6 +567,8 @@ async fn fill_stream(
     bound: Arc<BoundExpression>,
     column: &str,
     output_is_blob: bool,
+    recompute: bool,
+    gained: Arc<AtomicU64>,
 ) -> Result<impl Stream<Item = lance_core::Result<RecordBatch>> + Send + use<>> {
     let mut projection: Vec<String> = bound.roots.clone();
     projection.push(column.to_string());
@@ -521,14 +618,20 @@ async fn fill_stream(
             .column_by_name(ROW_ID)
             .ok_or_else(|| missing(ROW_ID))?;
 
-        // Only an unfilled live row gains a value; a deleted row has a null
-        // row id and keeps its (null) slot.
-        let unfilled = arrow::compute::is_null(existing.as_ref())?;
+        // Only an unfilled live row gains a value, or every live row under a
+        // recompute; a deleted row has a null row id and keeps its (null) slot.
         let live = arrow::compute::is_not_null(row_ids.as_ref())?;
-        let fill = arrow::compute::and(&unfilled, &live)?;
+        let fill = if recompute {
+            live
+        } else {
+            let unfilled = arrow::compute::is_null(existing.as_ref())?;
+            arrow::compute::and(&unfilled, &live)?
+        };
         let keep = arrow::compute::not(&fill)?;
 
         let computed = evaluate(&bound, &evaluation_batch(&batch, &bound, Some(&keep))?)?;
+        let values = arrow::compute::and(&fill, &arrow::compute::is_not_null(&computed)?)?;
+        gained.fetch_add(values.true_count() as u64, Ordering::Relaxed);
         let merged = arrow_select::zip::zip(&fill, &computed, existing)?;
         let merged = if output_is_blob {
             blob_array_from_binary(&merged, projected.field(0))?
@@ -737,7 +840,8 @@ mod tests {
             .await
             .unwrap();
         assert_eq!(no_op.rows_assigned, 0);
-        assert_eq!(no_op.source_version, 3);
+        // The fill, then the stamp recording what it computed from.
+        assert_eq!(no_op.source_version, 4);
         assert_eq!(no_op.published_version, None);
     }
 
@@ -777,8 +881,8 @@ mod tests {
     }
 
     /// A row is filled only by gaining a value, so an expression yielding null
-    /// settles at once instead of re-selecting the same rows forever. Nothing
-    /// is staged, so the version does not move either.
+    /// settles at once instead of re-selecting the same rows forever: the
+    /// second refresh finds the fragment signed and moves nothing.
     #[tokio::test]
     async fn test_refresh_converges_on_a_null_result() {
         let table = table_with("refresh_null_result", vec![1, 2, 3]).await;
@@ -792,28 +896,145 @@ mod tests {
 
         let first = table.refresh_column("maybe").await.unwrap();
         assert_eq!(first.rows_filled, 0);
-        assert_eq!(first.version, declared);
+        assert!(first.version > declared);
         assert_eq!(read(&table, "maybe").await, vec![None, None, None]);
 
         let again = table.refresh_column("maybe").await.unwrap();
         assert_eq!(again.rows_filled, 0);
-        assert_eq!(again.version, declared);
+        assert_eq!(again.version, first.version);
     }
 
-    /// The contract's boundary: a filled fragment is not revisited, so
-    /// mutating an input leaves the value computed at fill time.
+    /// A filled row whose input moved is recomputed: the update rewrites
+    /// the row into a fragment the stamp never signed, and only that one.
     #[tokio::test]
-    async fn test_refresh_does_not_observe_input_mutation() {
-        let table = table_with("refresh_mutation", vec![1]).await;
+    async fn test_refresh_recomputes_a_row_whose_input_moved() {
+        let table = table_with("refresh_mutation", vec![1, 2]).await;
         declare_doubled(&table).await.unwrap();
+        append(&table, vec![5]).await;
         table.refresh_column("doubled").await.unwrap();
-        assert_eq!(read(&table, "doubled").await, vec![Some(2)]);
+        assert_eq!(
+            read(&table, "doubled").await,
+            vec![Some(2), Some(4), Some(10)]
+        );
 
-        table.update().column("x", "3").execute().await.unwrap();
+        table
+            .update()
+            .column("x", "7")
+            .only_if("x = 5")
+            .execute()
+            .await
+            .unwrap();
+
+        let again = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(again.rows_filled, 1);
+        assert_eq!(
+            read(&table, "doubled").await,
+            vec![Some(2), Some(4), Some(14)]
+        );
+
+        let settled = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(settled.rows_filled, 0);
+        assert_eq!(settled.version, again.version);
+    }
+
+    /// A deleted row is never computed and the rows that stay keep their
+    /// values: a delete recomputes nothing and stamps nothing.
+    #[tokio::test]
+    async fn test_a_delete_recomputes_nothing() {
+        let table = table_with("refresh_delete", vec![1, 2, 3]).await;
+        declare_doubled(&table).await.unwrap();
+        let filled = table.refresh_column("doubled").await.unwrap();
+
+        table.delete("x = 2").await.unwrap();
+        let deleted = table.version().await.unwrap();
 
         let again = table.refresh_column("doubled").await.unwrap();
         assert_eq!(again.rows_filled, 0);
-        assert_eq!(read(&table, "doubled").await, vec![Some(2)]);
+        assert_eq!(again.version, deleted);
+        assert!(deleted > filled.version);
+        assert_eq!(read(&table, "doubled").await, vec![Some(2), Some(6)]);
+    }
+
+    /// Compaction copies inputs unchanged, so a fragment it builds from
+    /// signed ones is fresh: the refresh recomputes nothing and only records
+    /// the new fragment.
+    #[tokio::test]
+    async fn test_a_compaction_of_signed_fragments_recomputes_nothing() {
+        let table = table_with("refresh_compact_signed", vec![1, 2]).await;
+        declare_doubled(&table).await.unwrap();
+        append(&table, vec![5]).await;
+        table.refresh_column("doubled").await.unwrap();
+
+        table
+            .optimize(crate::table::OptimizeAction::Compact {
+                options: crate::table::CompactionOptions::default(),
+                remap_options: None,
+            })
+            .await
+            .unwrap();
+        let compacted = table.version().await.unwrap();
+
+        let carried = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(carried.rows_filled, 0);
+        assert_eq!(carried.version, compacted + 1);
+        let settled = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(settled.version, carried.version);
+        assert_eq!(
+            read(&table, "doubled").await,
+            vec![Some(2), Some(4), Some(10)]
+        );
+    }
+
+    /// A column declared before signatures existed has no map. Its first
+    /// refresh keeps the null-fill contract and enrolls what it read from;
+    /// from then on a moved input is recomputed like any other.
+    #[tokio::test]
+    async fn test_an_unsigned_column_is_enrolled_by_its_first_refresh() {
+        let table = table_with("refresh_legacy", vec![1, 2]).await;
+        declare_doubled(&table).await.unwrap();
+        table.refresh_column("doubled").await.unwrap();
+        table
+            .update()
+            .column("x", "3")
+            .only_if("x = 1")
+            .execute()
+            .await
+            .unwrap();
+
+        let native = table.as_native().unwrap();
+        let mut dataset = (*native.dataset.get().await.unwrap()).clone();
+        let declaration = dataset
+            .schema()
+            .field("doubled")
+            .unwrap()
+            .metadata
+            .iter()
+            .filter(|(key, _)| !key.starts_with("computed_refresh."))
+            .map(|(key, value)| (key.clone(), value.clone()))
+            .collect::<Vec<_>>();
+        dataset
+            .update_field_metadata()
+            .replace("doubled", declaration)
+            .unwrap()
+            .await
+            .unwrap();
+        table.checkout_latest().await.unwrap();
+
+        // Null-fill only: the moved row keeps the value it was filled with.
+        let enrolled = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(enrolled.rows_filled, 0);
+        assert_eq!(read(&table, "doubled").await, vec![Some(2), Some(4)]);
+
+        table
+            .update()
+            .column("x", "5")
+            .only_if("x = 3")
+            .execute()
+            .await
+            .unwrap();
+        let again = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(again.rows_filled, 1);
+        assert_eq!(read(&table, "doubled").await, vec![Some(4), Some(10)]);
     }
 
     /// A row rewrite before the first refresh materializes the declared
@@ -831,11 +1052,11 @@ mod tests {
         assert_eq!(read(&table, "doubled").await, vec![Some(6)]);
     }
 
-    /// The contract holds row by row, not fragment by fragment: revisiting a
-    /// fragment to fill one row must not recompute a filled row sitting beside
-    /// it, even where the input behind it has since changed.
+    /// A fragment compacted out of one the stamp never signed cannot vouch
+    /// for any of its rows: every live row is recomputed, the moved one
+    /// included.
     #[tokio::test]
-    async fn test_refresh_does_not_recompute_a_filled_row_beside_an_unfilled_one() {
+    async fn test_a_compaction_of_an_unsigned_fragment_recomputes_it() {
         let table = table_with("refresh_mixed", vec![1, 2]).await;
         declare_doubled(&table).await.unwrap();
         table.refresh_column("doubled").await.unwrap();
@@ -857,18 +1078,70 @@ mod tests {
             .unwrap();
 
         let result = table.refresh_column("doubled").await.unwrap();
-        assert_eq!(result.rows_filled, 1);
-        // 2 is the mutated row keeping the value it was filled with, not 200.
+        assert_eq!(result.rows_filled, 3);
+        assert_eq!(
+            read(&table, "doubled").await,
+            vec![Some(4), Some(10), Some(200)]
+        );
+    }
+
+    /// The gate's reproducer: a raw lance append may carry a value for the
+    /// computed column. Compaction cannot certify it, so the product is
+    /// recomputed and the supplied value replaced.
+    #[tokio::test]
+    async fn test_raw_append_values_are_not_trusted_after_compaction() {
+        use arrow_array::RecordBatchIterator;
+        use lance::Dataset;
+        use lance::dataset::{WriteMode, WriteParams};
+
+        let dir = tempfile::tempdir().unwrap();
+        let conn = connect(dir.path().to_str().unwrap())
+            .execute()
+            .await
+            .unwrap();
+        let batch = record_batch!(("x", Int32, [1, 2])).unwrap();
+        let table = conn
+            .create_table("raw_append", batch)
+            .execute()
+            .await
+            .unwrap();
+        declare_doubled(&table).await.unwrap();
+        table.refresh_column("doubled").await.unwrap();
+
+        let batch = record_batch!(("x", Int32, [5]), ("doubled", Int32, [Some(999_i32)])).unwrap();
+        let schema = batch.schema();
+        let uri = table.uri().await.unwrap();
+        Dataset::write(
+            RecordBatchIterator::new(vec![Ok(batch)], schema),
+            &uri,
+            Some(WriteParams {
+                mode: WriteMode::Append,
+                ..Default::default()
+            }),
+        )
+        .await
+        .unwrap();
+        table.checkout_latest().await.unwrap();
+        table
+            .optimize(crate::table::OptimizeAction::Compact {
+                options: crate::table::CompactionOptions::default(),
+                remap_options: None,
+            })
+            .await
+            .unwrap();
+
+        let result = table.refresh_column("doubled").await.unwrap();
+        assert_eq!(result.rows_filled, 3);
         assert_eq!(
             read(&table, "doubled").await,
             vec![Some(2), Some(4), Some(10)]
         );
     }
 
-    /// Filling a fragment must not disturb the values it already holds, which
-    /// is what makes a compaction-mixed fragment safe to revisit.
+    /// An appended fragment holds no values, so compacting it into a signed
+    /// one leaves the product fresh: only the appended rows are filled.
     #[tokio::test]
-    async fn test_refresh_preserves_already_filled_rows() {
+    async fn test_a_compaction_with_an_appended_fragment_fills_only_its_rows() {
         let table = table_with("refresh_preserves", vec![1, 2]).await;
         declare_doubled(&table).await.unwrap();
         table.refresh_column("doubled").await.unwrap();
@@ -995,7 +1268,8 @@ mod tests {
         assert_eq!(result.rows_failed, 0);
         assert_eq!(result.rows_remaining, 0);
         assert_eq!(result.source_version, 2);
-        assert_eq!(result.published_version, Some(3));
+        // The fill lands at 3; the stamp recording its inputs is published at 4.
+        assert_eq!(result.published_version, Some(4));
         assert_eq!(job.status().await.unwrap(), "finished");
         assert_eq!(
             read(&table, "doubled").await,
@@ -1059,31 +1333,32 @@ mod tests {
         assert_eq!(read(&table, "quotient").await, vec![Some(10)]);
     }
 
-    /// The gate's reproducer: an already-filled row's value must not be
-    /// re-evaluated either -- its input may have mutated into one the
-    /// expression chokes on.
+    /// A filled row whose input moved is re-evaluated, and a row whose
+    /// input did not move is not: the untouched fragment is never read, so
+    /// its poison input is never reached.
     #[tokio::test]
-    async fn test_a_filled_rows_value_is_never_evaluated() {
-        let table = table_with("refresh_filled_poison", vec![1, 2]).await;
+    async fn test_only_a_moved_rows_value_is_re_evaluated() {
+        let table = table_with("refresh_filled_poison", vec![1, 0]).await;
         table
             .add_columns()
-            .computed("quotient", "10 / x")
+            .computed("quotient", "10 / coalesce(nullif(x, 0), 1)")
             .execute()
             .await
             .unwrap();
         table.refresh_column("quotient").await.unwrap();
+        assert_eq!(read(&table, "quotient").await, vec![Some(10), Some(10)]);
 
+        append(&table, vec![5]).await;
         table
             .update()
-            .column("x", "0")
+            .column("x", "2")
             .only_if("x = 1")
             .execute()
             .await
             .unwrap();
-        append(&table, vec![5]).await;
 
         let result = table.refresh_column("quotient").await.unwrap();
-        assert_eq!(result.rows_filled, 1);
+        assert_eq!(result.rows_filled, 2);
         assert_eq!(
             read(&table, "quotient").await,
             vec![Some(2), Some(5), Some(10)]