diff --git a/scripts/bench/temp_calibrate.py b/scripts/bench/temp_calibrate.py
new file mode 100644
index 0000000..0dbb563
--- /dev/null
+++ b/scripts/bench/temp_calibrate.py
@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+"""Coolant temperature calibration helper (runs on Windows, reads the
+board over ssh).
+
+The raw->Celsius conversion in UAPI.md (C = raw * -0.09653 + 94) is an
+unverified best guess. This tool collects reference points - a measured
+real temperature paired with the machine's raw ADC readings - and fits
+the actual line for THIS machine.
+
+Usage:
+  temp_calibrate.py watch                 live raw + current-formula C
+  temp_calibrate.py point <measured_C>    record a calibration point
+  temp_calibrate.py fit                   fit and print the calibration
+
+Points accumulate in temp_calibration.json next to this script. Take at
+least two points as far apart in temperature as practical (e.g. cold
+machine in the morning, and warm after a fan-off soak with the flow
+heater on).
+"""
+import json
+import os
+import subprocess
+import sys
+import time
+
+HOST = '172.16.1.97'
+STORE = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'temp_calibration.json')
+
+# Current unverified guess, for comparison only.
+GUESS_SLOPE, GUESS_OFFSET = -0.09653, 94.0
+
+
+def board(cmd):
+    r = subprocess.run(['wsl', '-d', 'forge-yocto', '--', 'ssh',
+                        '-o', 'PreferredAuthentications=none',
+                        'root@' + HOST, cmd],
+                       capture_output=True, text=True, timeout=30)
+    return r.stdout.strip()
+
+
+def raws(samples=5, delay=1.0):
+    """Average several readings of both sensors (ADC noise is real)."""
+    acc1, acc2, n = 0, 0, 0
+    for _ in range(samples):
+        out = board('cat /sys/glowforge/pic/water_temp_1 /sys/glowforge/pic/water_temp_2').split()
+        if len(out) == 2:
+            acc1 += int(out[0]); acc2 += int(out[1]); n += 1
+        time.sleep(delay)
+    return (acc1 / n, acc2 / n) if n else (None, None)
+
+
+def load():
+    if os.path.exists(STORE):
+        with open(STORE) as f:
+            return json.load(f)
+    return {'points': []}
+
+
+def save(data):
+    with open(STORE, 'w') as f:
+        json.dump(data, f, indent=2)
+
+
+def fit(points, key):
+    """Least-squares line: measured_C = slope * raw + offset."""
+    xs = [p[key] for p in points]
+    ys = [p['measured_c'] for p in points]
+    n = len(xs)
+    mx = sum(xs) / n
+    my = sum(ys) / n
+    den = sum((x - mx) ** 2 for x in xs)
+    if den == 0:
+        return None, None
+    slope = sum((x - mx) * (y - my) for x, y in zip(xs, ys)) / den
+    return slope, my - slope * mx
+
+
+def main():
+    mode = sys.argv[1] if len(sys.argv) > 1 else 'watch'
+
+    if mode == 'watch':
+        print('raw1(down) raw2(up)   guess-C down/up   (ctrl-C to stop)')
+        while True:
+            r1, r2 = raws(1, 0)
+            print('  %6.1f   %6.1f      %.2f / %.2f'
+                  % (r1, r2, r1 * GUESS_SLOPE + GUESS_OFFSET,
+                     r2 * GUESS_SLOPE + GUESS_OFFSET))
+            time.sleep(2)
+
+    elif mode == 'point':
+        measured = float(sys.argv[2])
+        note = sys.argv[3] if len(sys.argv) > 3 else ''
+        print('sampling raws (10 s)...')
+        r1, r2 = raws()
+        data = load()
+        data['points'].append({'measured_c': measured, 'raw1': r1, 'raw2': r2,
+                               'note': note, 'when': time.strftime('%Y-%m-%d %H:%M:%S')})
+        save(data)
+        print('recorded: measured %.2f C  raw1=%.1f raw2=%.1f  (%d points total)'
+              % (measured, r1, r2, len(data['points'])))
+
+    elif mode == 'fit':
+        data = load()
+        pts = data['points']
+        if len(pts) < 2:
+            print('need at least 2 points (have %d)' % len(pts))
+            return 1
+        print('points:')
+        for p in pts:
+            print('  %6.2f C   raw1=%.1f raw2=%.1f   %s %s'
+                  % (p['measured_c'], p['raw1'], p['raw2'], p['when'], p['note']))
+        span = max(p['measured_c'] for p in pts) - min(p['measured_c'] for p in pts)
+        print('\ntemperature span: %.2f C%s' % (span, '  (WARNING: <3 C span, fit is weak)' if span < 3 else ''))
+        for key, label in (('raw1', 'downstream (water_temp_1)'), ('raw2', 'upstream (water_temp_2)')):
+            slope, offset = fit(pts, key)
+            print('\n%s:' % label)
+            print('  fitted:  C = raw * %.6f + %.4f' % (slope, offset))
+            print('  guess:   C = raw * %.6f + %.4f' % (GUESS_SLOPE, GUESS_OFFSET))
+            for raw in (600, 650, 700, 750, 800):
+                print('    raw %3d -> fitted %6.2f C   guess %6.2f C   diff %+.2f'
+                      % (raw, raw * slope + offset, raw * GUESS_SLOPE + GUESS_OFFSET,
+                         (raw * slope + offset) - (raw * GUESS_SLOPE + GUESS_OFFSET)))
+    else:
+        print(__doc__)
+        return 1
+    return 0
+
+
+if __name__ == '__main__':
+    sys.exit(main())