diff --git a/ui/src/views/Forecast.jsx b/ui/src/views/Forecast.jsx
index 5982bfd..3f4fb3e 100644
--- a/ui/src/views/Forecast.jsx
+++ b/ui/src/views/Forecast.jsx
@@ -419,22 +419,31 @@ export default function Forecast({ sources = [], sourceId, versions = [], versio
     const dateNames = new Set(colMetaRef.current.filter(c => c.role === 'date').map(c => c.cname))
     const ITER_ORDER = ['baseline', 'scale', 'recode', 'clone']
 
+    // A slice carries every value as a string -- it is built from filters the
+    // grid reports and from a payload that has to survive JSON. Perspective
+    // matches on type, so a string '2027' against an integer column does not
+    // filter to nothing, it is dropped: the ledger then totalled rows the pivot
+    // was hiding, which is how a season filter on sseas_e went unnoticed while
+    // the numbers disagreed by exactly the out-of-season rows.
+    const schema = await tableRef.current.schema()
+    const typed = (col, val) => {
+      switch (schema[col]) {
+        case 'integer': case 'float': return Number(val)
+        case 'boolean': return val === true || val === 'true'
+        case 'date': case 'datetime': return Number(val)
+        default: return String(val)
+      }
+    }
+
     async function totalsFor(sliceObj) {
-      const filters = [
-        // pf_segment and pf_bucket are computed server-side but are ordinary
-        // string columns in the loaded table, so here they filter directly. They
-        // have to be applied, or the ledger totals a wider selection than the
-        // operation will write.
-        ...Object.entries(sliceObj)
-          .filter(([col]) => COMPUTED_SLICE_COLS.has(col))
-          .map(([col, val]) => [col, '==', String(val)]),
-        ...Object.entries(sliceObj)
-          .filter(([col]) => dimNames.has(col))
-          .map(([col, val]) => [col, '==', val]),
-        ...Object.entries(sliceObj)
-          .filter(([col]) => dateNames.has(col))
-          .map(([col, val]) => [col, '==', Number(val)]),
-      ]
+      // pf_segment and pf_bucket are computed server-side but are ordinary
+      // columns in the loaded table, so here they filter directly. They have to
+      // be applied, or the ledger totals a wider selection than the operation
+      // will write.
+      const filters = Object.entries(sliceObj)
+        .filter(([col]) => COMPUTED_SLICE_COLS.has(col) || dimNames.has(col) || dateNames.has(col)
+                        || schema[col] !== undefined)
+        .map(([col, val]) => [col, '==', typed(col, val)])
       const view = await tableRef.current.view({ filter: filters })
       const rows = await view.to_json()
       await view.delete()