diff --git a/.github/workflows/wheels.yml b/.github/workflows/wheels.yml
index e4d62e0..7a11149 100644
--- a/.github/workflows/wheels.yml
+++ b/.github/workflows/wheels.yml
@@ -5,6 +5,11 @@ on:
     tags: ["v*"]
   workflow_dispatch:
 
+# Default the GITHUB_TOKEN to read-only; the publish job opts into
+# id-token: write for PyPI OIDC and nothing else needs write.
+permissions:
+  contents: read
+
 # Builds the Python wheels (`pip install sqlite-predict`). Each wheel bundles the
 # zero-dependency loadable compiled for its platform. Staging runs `make
 # python-src` on the host to drop the amalgamation + SQLite ext headers into the
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7be2549..5b02e38 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,6 +6,16 @@ project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Fixed
+
+- **Forest prediction accumulates the learning-rate product in double**, matching
+  the trainer. The serving path multiplied the learning rate by a tree value as
+  `float * float` before widening to the `double` accumulator, which could
+  overflow to infinity (and, under x87 excess precision, differ by platform)
+  before the widening. The product and the forest accumulation are now computed in
+  double (individual tree values are still `f32`); served predictions may shift in
+  their least-significant digits.
+
 ## [0.2.0] - 2026-07-31
 
 ### Changed
diff --git a/predict-student.c b/predict-student.c
index ff342fe..1946bfd 100644
--- a/predict-student.c
+++ b/predict-student.c
@@ -276,7 +276,10 @@ int predict0_forest_predict_row(const Forest *f, const f32 *x, f64 *scbuf, char
   if (f->task == 1) {
     f64 s = f->init[0];
     for (int j = 0; j < f->n_trees; j++)
-      s += f->lr * forest_tree_value(f, j, x);
+      /* (f64) before the multiply: keep the product in double, matching the
+       * trainer (predict-train.c), so a float*float never overflows or picks up
+       * platform-dependent excess precision before it reaches the accumulator. */
+      s += (f64)f->lr * forest_tree_value(f, j, x);
     *has_conf = 0;
     *pred = sqlite3_mprintf("%.17g", s);
     return *pred ? SQLITE_OK : SQLITE_NOMEM;
@@ -285,7 +288,8 @@ int predict0_forest_predict_row(const Forest *f, const f32 *x, f64 *scbuf, char
     scbuf[c] = f->init[c];
   for (int r = 0; r < f->n_rounds; r++)
     for (int s = 0; s < f->n_score; s++)
-      scbuf[s] += f->lr * forest_tree_value(f, r * f->n_score + s, x);
+      /* (f64) before the multiply, as above and in the trainer */
+      scbuf[s] += (f64)f->lr * forest_tree_value(f, r * f->n_score + s, x);
   f64 mx = scbuf[0];
   int arg = 0;
   for (int c = 1; c < f->n_score; c++)