diff --git a/rigforge.sh b/rigforge.sh
index 2fbb542..df43a87 100755
--- a/rigforge.sh
+++ b/rigforge.sh
@@ -2123,7 +2123,7 @@ _xmrig_bench() { # <bin> <bench-size> <config|"">
     # The MEDIAN clock over the window — robust to the brief low-clock dataset-init phase, which the raw
     # min would mistake for thermal throttling (#62).
     # shellcheck disable=SC2086 # $freqs is an intentional word-split list of samples
-    if [ -n "${BENCH_FREQ_FILE:-}" ]; then printf '%s' "$(_median $freqs)" >"$BENCH_FREQ_FILE"; fi
+    if [ -n "${BENCH_FREQ_FILE:-}" ]; then _median $freqs | awk '{printf "%d", $0+0}' >"$BENCH_FREQ_FILE"; fi
     cat "$logf" "$out" 2>/dev/null
     rm -f "$out" "$logf" "$tmpcfg" 2>/dev/null || true
 }
@@ -2207,9 +2207,9 @@ _measure() { # <prefetch> <yield> <threads> <onegb> <priority> <hpjit> <cacheqos
             [ -n "$s" ] || s=0
             samples+=("$s")
             bf=$(cat "$TUNE_TMP/freq" 2>/dev/null)
-            # _median emits a fractional kHz for an even sample count (e.g. 4627500.5); floor it to whole kHz
-            # so the integer guard below keeps the reading instead of dropping it — a dropped reading would
-            # leave min_freq_mhz null, silently disabling this candidate's #62 throttle check.
+            # The freq writer above already floors to whole kHz (awk's OFMT prints large medians in
+            # scientific notation, e.g. 4.6275e+06, which this floor would otherwise mangle into "4").
+            # This is now just a belt-and-suspenders trim of a stray fractional part.
             bf=${bf%.*}
             case "$bf" in '' | *[!0-9]*) ;; *) if [ -z "$minfk" ] || [ "$bf" -lt "$minfk" ]; then minfk="$bf"; fi ;; esac
             wv=$(cat "$TUNE_TMP/watts" 2>/dev/null)
diff --git a/tests/run.sh b/tests/run.sh
index f09be45..debd1b2 100644
--- a/tests/run.sh
+++ b/tests/run.sh
@@ -4295,6 +4295,48 @@ out="$(confirm_run 'c=$(cat "$ACTR" 2>/dev/null||echo 0);c=$((c+1));echo $c>"$AC
 assert_contains "confirm reverts a live regression (#64)" "$out" "Reverted:"
 assert_eq "reverted -> previous (none) restored (#64)" "$([ -f "$OVR" ] && echo present || echo gone)" "gone"
 
+# #266: unit-style — an even sample count whose middle pair sums to an ODD kHz total makes awk print the
+# median in scientific notation (e.g. 4.6275e+06 for 4627000/4628001 kHz, real DVFS jitter between reads).
+# The freq writer at rigforge.sh:2120 must floor that back to whole kHz, not leave it for the consumer's
+# `.`-floor to truncate into a stray "4" that reads as 0 MHz and false-trips the #62 throttle guard.
+echo "== unit: tune bench freq writer floors a sci-notation median to whole kHz (#266) =="
+FQDONE="$SANDBOX/freq266.done"
+export FQDONE
+FQBIN="$SANDBOX/freq266-xmrig"
+# The fake prints 'benchmark finished' once released: the bench poll loop greps for that line and breaks,
+# so no extra clock sample sneaks in after the replayed ones (a just-exited fake can linger as a zombie
+# that `kill -0` still sees, which would otherwise buy the loop one more iteration).
+cat >"$FQBIN" <<'EOF'
+#!/usr/bin/env bash
+echo "speed 1000.0 H/s max 1000.0 H/s"
+for _ in $(seq 1 500); do [ -f "$FQDONE" ] && break; sleep 0.02; done # bounded: never outlive the test
+echo "benchmark finished"
+EOF
+chmod +x "$FQBIN"
+freqwrite() { ( # <freq sample kHz>... -> BENCH_FREQ_FILE content, via the real _xmrig_bench freq-write path
+    source "$SCRIPT"
+    local vals=("$@") n=$# fqout="$SANDBOX/freq266.out" ctr="$SANDBOX/freq266.ctr"
+    rm -f "$FQDONE" "$fqout"
+    printf 0 >"$ctr"
+    # Deterministic clock reader: replay exactly the given samples, one per call, then signal the fake
+    # xmrig to finish and go silent — no reliance on real hardware jitter or wall-clock timing.
+    # The call counter lives in a FILE: each sample is read via $(...), a subshell where a plain shell
+    # variable increment would not persist. `return 0` keeps set -e happy on the non-final sample.
+    _cpu_eff_khz() {
+        local i
+        i=$(cat "$ctr")
+        if [ "$i" -ge "$n" ]; then return 0; fi # exhausted -> empty reading, which the sampler skips
+        printf '%s' "${vals[$i]}"
+        printf '%s' "$((i + 1))" >"$ctr"
+        if [ "$((i + 1))" -ge "$n" ]; then touch "$FQDONE"; fi
+        return 0
+    }
+    BENCH_FREQ_FILE="$fqout" _xmrig_bench "$FQBIN" 1M "" >/dev/null
+    cat "$fqout"
+); }
+assert_eq "bench freq median floors a sci-notation kHz average to whole kHz, not a stray digit (#266)" \
+    "$(freqwrite 4627000 4628001)" "4627500"
+
 # #62: thermal-throttle rejection. A LOW clock source makes every candidate's window "throttled" — a
 # faster-but-throttled candidate must NOT be adopted (its number reflects the throttle, not the config),
 # and the throttle must be recorded in the log. With TUNE_MIN_FREQ_MHZ=0 the skip is disabled.
@@ -4329,6 +4371,39 @@ assert_eq "throttled faster candidate not adopted (#62)" "$(J "$OVR" '.randomx.s
 out="$(throttle_run 0)"
 assert_eq "TUNE_MIN_FREQ_MHZ=0 disables the throttle skip (#62 control)" "$(J "$OVR" '.randomx.scratchpad_prefetch_mode')" "2"
 
+# #266: regression — a HEALTHY (~4.6 GHz) candidate whose clock jitters by 1 kHz between two reads gets a
+# fractional kHz median (4627500.5) that awk printed as 4.6275e+06; the consumer's `.`-floor mangled that
+# into "4" -> 0 MHz -> the #62 guard falsely flagged the candidate as throttled and refused to adopt it.
+# scaling_cur_freq is a FIFO fed exactly one odd-sum pair per bench window (reads block until fed), so the
+# window always sees an even sample count with a fractional median — no reliance on poll-loop timing.
+echo "== black-box: tune bench freq median doesn't false-trip the #62 throttle guard (#266) =="
+mkdir -p "$TN/cpu266/cpu0/cpufreq"
+FF266="$TN/cpu266/cpu0/cpufreq/scaling_cur_freq"
+DONE266="$TN/done266"
+rm -f "$FF266" "$DONE266"
+mkfifo "$FF266"
+# Feeder: serve the pair, then release the fake xmrig so the bench window closes after exactly 2 samples.
+(while :; do printf '4627000\n' >"$FF266" && printf '4628001\n' >"$FF266" && touch "$DONE266" || exit; done) &
+FEED266=$!
+cat >"$BD/xmrig" <<EOF
+#!/usr/bin/env bash
+echo "speed 1000.0 H/s max 1000.0 H/s"
+for _ in \$(seq 1 500); do [ -f "$DONE266" ] && break; sleep 0.02; done # bounded wait
+rm -f "$DONE266"
+echo "benchmark finished" # breaks the bench poll loop BEFORE a zombie-pid extra iteration samples a 3rd clock
+EOF
+chmod +x "$BD/xmrig"
+out="$(cd "$TN" && PATH="$STUBS:$PATH" CPU_SYSFS="$TN/cpu266" CPUFREQ_MAX="$TN/cpu_max" TUNE_MIN_FREQ_MHZ=4000 \
+    TUNE_ITERS=1 TUNE_SEEDS=auto TUNE_PREFETCH_MODES=1 TUNE_YIELDS=false TUNE_THREADS=-1 \
+    RIGFORGE_HOME="$PWD" bash "$SCRIPT" tune </dev/null 2>&1)"
+rc=$?
+kill "$FEED266" 2>/dev/null
+wait "$FEED266" 2>/dev/null
+rm -f "$FF266" "$DONE266"
+assert_rc "healthy fractional-median tune exits 0 (#266)" "$rc" "0"
+assert_absent "healthy candidate NOT flagged as throttled (#266)" "$out" "throttled to"
+assert_eq "healthy candidate recorded as not throttled, eligible for adoption (#266)" "$(J "$TLOG" '.results[0].throttled')" "false"
+
 # #265: _seed_wr / _seed_g must keep an explicit base-config false instead of jq `//` flipping it to
 # the true default; with the key absent, the true default still applies.
 echo "== unit: tune seeds keep explicit false (#265) =="