diff --git a/changelog.json b/changelog.json
index e51a4a8..9ac85a2 100644
--- a/changelog.json
+++ b/changelog.json
@@ -3,10 +3,14 @@
     "version": "0.22.3",
     "date": "",
     "en": [
-      "The Kenwood backend is now exercised against a rig that answers: OpsLog talks to the TS-2000 emulator it already carries for ACOM amplifiers, so frequency, mode, VFO, split and PTT are checked end to end. Testing on a real Kenwood is still needed, but the dialogue itself is no longer untried."
+      "The Kenwood backend is now exercised against a rig that answers: OpsLog talks to the TS-2000 emulator it already carries for ACOM amplifiers, so frequency, mode, VFO, split and PTT are checked end to end. Testing on a real Kenwood is still needed, but the dialogue itself is no longer untried.",
+      "Kenwood: split is now detected from the transmit and receive VFO (FR/FT) when the radio leaves it out of its status frame — the case seen on a Flex in Kenwood CAT mode, where the frequency read correctly but split never appeared. Radios that report it in the status frame are unaffected.",
+      "The CAT protocol trace covers the Kenwood backend as well, not just CI-V. Settings → CAT."
     ],
     "fr": [
-      "Le backend Kenwood est désormais éprouvé face à une radio qui répond : OpsLog dialogue avec l'émulateur TS-2000 qu'il embarque déjà pour les amplis ACOM, ce qui vérifie fréquence, mode, VFO, split et PTT de bout en bout. Un essai sur un vrai Kenwood reste nécessaire, mais le dialogue n'est plus non testé."
+      "Le backend Kenwood est désormais éprouvé face à une radio qui répond : OpsLog dialogue avec l'émulateur TS-2000 qu'il embarque déjà pour les amplis ACOM, ce qui vérifie fréquence, mode, VFO, split et PTT de bout en bout. Un essai sur un vrai Kenwood reste nécessaire, mais le dialogue n'est plus non testé.",
+      "Kenwood : le split est désormais détecté à partir des VFO d'émission et de réception (FR/FT) quand la radio ne le renseigne pas dans sa trame d'état — le cas observé sur un Flex en mode CAT Kenwood, où la fréquence était juste mais le split n'apparaissait jamais. Les radios qui le signalent normalement ne changent pas.",
+      "La trace du protocole CAT couvre aussi le backend Kenwood, plus seulement le CI-V. Réglages → CAT."
     ]
   },
   {
diff --git a/frontend/src/components/SettingsModal.tsx b/frontend/src/components/SettingsModal.tsx
index c5f0a60..9a3d9ba 100644
--- a/frontend/src/components/SettingsModal.tsx
+++ b/frontend/src/components/SettingsModal.tsx
@@ -2516,7 +2516,7 @@ export function SettingsModal({ onClose, onSaved, initialSection, onMainPaneChan
               </div>
             </>
             )}
-            {(catCfg.backend === 'icom' || catCfg.backend === 'xiegu') && (
+            {['icom', 'xiegu', 'kenwood'].includes(catCfg.backend) && (
               <div className="border-t border-border/60 pt-3">
                 <label className="flex items-center gap-2 text-sm cursor-pointer">
                   <Checkbox checked={civTrace} onCheckedChange={(c) => { setCivTrace(!!c); SetCIVTrace(!!c); }} />
diff --git a/frontend/src/lib/i18n.tsx b/frontend/src/lib/i18n.tsx
index 78ab513..9505f3f 100644
--- a/frontend/src/lib/i18n.tsx
+++ b/frontend/src/lib/i18n.tsx
@@ -277,7 +277,7 @@ const en: Dict = {
   'extsvc.hint': 'Upload logged QSOs to online logbooks. Each service uploads automatically on a new QSO when enabled; timing is per-service (immediate, or a 1–2 min delay so a mis-logged QSO can still be fixed first).',
   'hw.motorTxInhibit': 'Inhibit transmission while the antenna is moving', 'hw.motorTxInhibitHint': 'Blocks the FlexRadio from transmitting while the elements move (needs FlexRadio in API mode + this antenna enabled). No effect with other radios.', 'hw.motorAntenna': 'Ultrabeam / Steppir', 'hw.motorEnable': 'Enable antenna control', 'hw.motorType': 'Antenna type', 'hw.motorTransport': 'Connection', 'hw.motorTcp': 'Network (TCP)', 'hw.motorSerial': 'Serial (COM)', 'hw.motorCom': 'Serial port', 'hw.motorBaud': 'Baud', 'hw.steppirHint': 'SteppIR controllers are RS-232 serial (the DATA OUT DB9 port). Serial = a USB↔RS-232 (FTDI) adapter, shown as a COM port. Network = a serial-to-Ethernet bridge (as for the Ultrabeam).', 'hw.steppirRange': 'Tunable range', 'hw.steppirRangeHint': "The SteppIR's frequency coverage. On a band outside this range (e.g. 30 m on a 20 m–6 m SteppIR) OpsLog won't try to tune the antenna and won't inhibit transmission. Default 13–54 MHz (20 m–6 m); widen the low edge (e.g. 6) for a 40 m-equipped SteppIR.", 'hw.ultrabeam': 'Antenna (Ultrabeam)', 'hw.audioVoice': 'Audio devices & voice keyer',
   // CAT panel body
-  'cat.enable': 'Enable CAT', 'cat.backend': 'Backend', 'cat.optOmnirig': 'OmniRig (any rig, Windows COM)', 'cat.optFlex': 'FlexRadio / SmartSDR (native)', 'cat.share': 'Share CAT with other programs', 'cat.shareHint': 'A native CAT backend owns the radio serial port, so no other program can reach it. This lets WSJT-X, JTDX, MSHV or Log4OM talk to the rig THROUGH OpsLog: in the other program pick the rig model "Hamlib NET rigctl" and enter 127.0.0.1:4532. Works with every backend, not only the native ones.', 'cat.sharePort': 'Sharing port', 'cat.optXiegu': 'Xiegu (native CI-V)', 'cat.xieguPort': 'Xiegu CAT port', 'cat.xieguBaudHint': 'Must match the radio menu (G90 default: 19200).', 'cat.xieguAddrHint': 'Factory CI-V address of the G90/X6100 family: 0x70.', 'cat.optYaesu': 'Yaesu (native CAT)', 'cat.optKenwood': 'Kenwood (native)', 'cat.civTrace': 'Log the CI-V protocol', 'cat.civTraceHint': 'Writes every CI-V frame to and from the radio into the log, as hex. For reporting a rig that answers oddly \u2014 a button that does the wrong thing, a frequency that jumps. Session only: it is a diagnostic, and it makes the log large.', 'cat.kenwoodPort': 'Kenwood COM port', 'cat.kenwoodPortHint': 'The rig\u2019s CAT/USB serial port (TS-590, TS-890, TS-990, TS-2000, and Elecraft K3/K4 which speak the same dialect).', 'cat.kenwoodBaudHint': 'Must match MENU on the radio: a TS-590 leaves the factory at 9600, a TS-890 at 115200.', 'cat.yaesuPort': 'Yaesu CAT port', 'cat.yaesuPortHint': 'The rig CAT port — on an FTDX10/FTDX101 over USB this is the ENHANCED COM port, not the standard one.', 'cat.yaesuBaudHint': 'Must match the radio menu (FTDX10/FTDX101 default: 38400).', 'cat.optIcom': 'Icom CI-V (USB serial)', 'cat.optIcomNet': 'Icom CI-V (network / remote)', 'cat.optTci': 'TCI (Expert Electronics / SunSDR)',
+  'cat.enable': 'Enable CAT', 'cat.backend': 'Backend', 'cat.optOmnirig': 'OmniRig (any rig, Windows COM)', 'cat.optFlex': 'FlexRadio / SmartSDR (native)', 'cat.share': 'Share CAT with other programs', 'cat.shareHint': 'A native CAT backend owns the radio serial port, so no other program can reach it. This lets WSJT-X, JTDX, MSHV or Log4OM talk to the rig THROUGH OpsLog: in the other program pick the rig model "Hamlib NET rigctl" and enter 127.0.0.1:4532. Works with every backend, not only the native ones.', 'cat.sharePort': 'Sharing port', 'cat.optXiegu': 'Xiegu (native CI-V)', 'cat.xieguPort': 'Xiegu CAT port', 'cat.xieguBaudHint': 'Must match the radio menu (G90 default: 19200).', 'cat.xieguAddrHint': 'Factory CI-V address of the G90/X6100 family: 0x70.', 'cat.optYaesu': 'Yaesu (native CAT)', 'cat.optKenwood': 'Kenwood (native)', 'cat.civTrace': 'Log the CAT protocol', 'cat.civTraceHint': 'Writes every CAT frame to and from the radio into the log \u2014 CI-V as hex, Kenwood as the text it exchanges. For reporting a rig that answers oddly \u2014 a button that does the wrong thing, a frequency that jumps. Session only: it is a diagnostic, and it makes the log large.', 'cat.kenwoodPort': 'Kenwood COM port', 'cat.kenwoodPortHint': 'The rig\u2019s CAT/USB serial port (TS-590, TS-890, TS-990, TS-2000, and Elecraft K3/K4 which speak the same dialect).', 'cat.kenwoodBaudHint': 'Must match MENU on the radio: a TS-590 leaves the factory at 9600, a TS-890 at 115200.', 'cat.yaesuPort': 'Yaesu CAT port', 'cat.yaesuPortHint': 'The rig CAT port — on an FTDX10/FTDX101 over USB this is the ENHANCED COM port, not the standard one.', 'cat.yaesuBaudHint': 'Must match the radio menu (FTDX10/FTDX101 default: 38400).', 'cat.optIcom': 'Icom CI-V (USB serial)', 'cat.optIcomNet': 'Icom CI-V (network / remote)', 'cat.optTci': 'TCI (Expert Electronics / SunSDR)',
   'cat.icomNetHost': 'Rig IP / hostname', 'cat.icomNetUser': 'Network user (ID)', 'cat.icomNetPass': 'Network password',
   'cat.icomNetHint': "Connects to the rig's built-in LAN server directly — no RS-BA1 or Remote Utility needed (close them first). Use the Network User1 ID/Password set in the rig's Network menu. A rig in standby is powered on automatically.",
   'cat.icomNetAudio': 'Stream RX audio over the network (experimental)',
@@ -678,7 +678,7 @@ const fr: Dict = {
   'rot.enable': 'Activer le contrôle du rotator', 'rot.testOkRG': 'Connecté — le Rotator Genius a accepté la commande (rotation vers 0°).',  'rot.type': 'Type de rotator', 'rot.rotatorNum': 'Rotator n°', 'rot.rgHint': 'Parle directement à un Rotator Genius 4O3A en TCP (port 9006 par défaut) — sans PstRotator. Le n° choisit lequel des deux rotators du boîtier piloter.', 'rot.arcoHint': "Parle directement à tout contrôleur réglé sur Yaesu GS-232A — sans PstRotator. microHAM ARCO : régler Config → LAN → CONTROL PROTOCOL (ou USB CONTROL PROTOCOL) sur « Yaesu GS-232A » ; en USB la vitesse est sans importance. ERC (Easy Rotor Control, ERC Mini compris) : son émulation DOIT être réglée sur GS-232 — un ERC laissé en Hy-Gain DCU-1 parle un autre jeu de commandes et ne répondra pas — puis choisir son port COM et la même vitesse que dans la configuration de l'ERC.", 'rot.hint': "OpsLog envoie des commandes UDP à PstRotator. Active l'écouteur UDP de PstRotator (Setup → Communication → UDP) avant de tester.",
   'extsvc.hint': "Envoie les QSO enregistrés vers des carnets en ligne. Chaque service upload automatiquement à chaque nouveau QSO si activé ; le délai est propre à chaque service (immédiat, ou 1–2 min pour corriger un QSO mal saisi avant).",
   'hw.motorTxInhibit': "Inhiber la transmission pendant que l'antenne bouge", 'hw.motorTxInhibitHint': "Empêche le FlexRadio d'émettre pendant que les éléments bougent (nécessite le FlexRadio en API + cette antenne activée). Sans effet avec les autres radios.", 'hw.motorAntenna': 'Antenne motorisée', 'hw.motorEnable': "Activer le contrôle de l'antenne", 'hw.motorType': "Type d'antenne", 'hw.motorTransport': 'Connexion', 'hw.motorTcp': 'Réseau (TCP)', 'hw.motorSerial': 'Série (COM)', 'hw.motorCom': 'Port série', 'hw.motorBaud': 'Débit', 'hw.steppirHint': "Les contrôleurs SteppIR sont en RS-232 série (port DB9 « DATA OUT »). Série = un adaptateur USB↔RS-232 (FTDI), vu comme un port COM. Réseau = un pont série-Ethernet (comme pour l'Ultrabeam).", 'hw.steppirRange': 'Plage accordable', 'hw.steppirRangeHint': "La couverture en fréquence de la SteppIR. Sur une bande hors de cette plage (p. ex. 30 m avec une SteppIR 20 m-6 m), OpsLog n'essaie pas d'accorder l'antenne et n'inhibe pas l'émission. Défaut 13-54 MHz (20 m-6 m) ; abaisse la borne basse (p. ex. 6) pour une SteppIR équipée 40 m.", 'hw.ultrabeam': 'Antenne (Ultrabeam)', 'hw.audioVoice': 'Périphériques audio & manipulateur vocal',
-  'cat.enable': 'Activer le CAT', 'cat.backend': 'Backend', 'cat.optOmnirig': 'OmniRig (tout poste, COM Windows)', 'cat.optFlex': 'FlexRadio / SmartSDR (natif)', 'cat.share': 'Partager le CAT avec les autres logiciels', 'cat.shareHint': "Un backend CAT natif occupe le port série de la radio, aucun autre logiciel ne peut donc y accéder. Ceci permet à WSJT-X, JTDX, MSHV ou Log4OM de dialoguer avec la radio À TRAVERS OpsLog : dans l'autre logiciel, choisissez le modèle « Hamlib NET rigctl » et saisissez 127.0.0.1:4532. Fonctionne avec tous les backends, pas seulement les natifs.", 'cat.sharePort': 'Port de partage', 'cat.optXiegu': 'Xiegu (CI-V natif)', 'cat.xieguPort': 'Port CAT Xiegu', 'cat.xieguBaudHint': 'Doit correspondre au menu de la radio (G90 par défaut : 19200).', 'cat.xieguAddrHint': "Adresse CI-V d'usine de la famille G90/X6100 : 0x70.", 'cat.optYaesu': 'Yaesu (CAT natif)', 'cat.optKenwood': 'Kenwood (natif)', 'cat.civTrace': 'Journaliser le protocole CI-V', 'cat.civTraceHint': '\u00c9crit dans le journal chaque trame CI-V \u00e9chang\u00e9e avec la radio, en hexad\u00e9cimal. Pour signaler une radio qui r\u00e9pond de travers \u2014 un bouton qui fait autre chose, une fr\u00e9quence qui saute. Valable pour la session seulement : c\u2019est un diagnostic, et le journal grossit vite.', 'cat.kenwoodPort': 'Port COM Kenwood', 'cat.kenwoodPortHint': 'Le port s\u00e9rie CAT/USB de la radio (TS-590, TS-890, TS-990, TS-2000, ainsi que les Elecraft K3/K4 qui parlent le m\u00eame dialecte).', 'cat.kenwoodBaudHint': 'Doit correspondre au MENU de la radio : un TS-590 sort d\u2019usine \u00e0 9600, un TS-890 \u00e0 115200.', 'cat.yaesuPort': 'Port CAT Yaesu', 'cat.yaesuPortHint': "Le port CAT de la radio — sur un FTDX10/FTDX101 en USB c'est le port COM ENHANCED, pas le standard.", 'cat.yaesuBaudHint': 'Doit correspondre au menu de la radio (FTDX10/FTDX101 par défaut : 38400).', 'cat.optIcom': 'Icom CI-V (USB série)', 'cat.optIcomNet': 'Icom CI-V (réseau / remote)', 'cat.optTci': 'TCI (Expert Electronics / SunSDR)',
+  'cat.enable': 'Activer le CAT', 'cat.backend': 'Backend', 'cat.optOmnirig': 'OmniRig (tout poste, COM Windows)', 'cat.optFlex': 'FlexRadio / SmartSDR (natif)', 'cat.share': 'Partager le CAT avec les autres logiciels', 'cat.shareHint': "Un backend CAT natif occupe le port série de la radio, aucun autre logiciel ne peut donc y accéder. Ceci permet à WSJT-X, JTDX, MSHV ou Log4OM de dialoguer avec la radio À TRAVERS OpsLog : dans l'autre logiciel, choisissez le modèle « Hamlib NET rigctl » et saisissez 127.0.0.1:4532. Fonctionne avec tous les backends, pas seulement les natifs.", 'cat.sharePort': 'Port de partage', 'cat.optXiegu': 'Xiegu (CI-V natif)', 'cat.xieguPort': 'Port CAT Xiegu', 'cat.xieguBaudHint': 'Doit correspondre au menu de la radio (G90 par défaut : 19200).', 'cat.xieguAddrHint': "Adresse CI-V d'usine de la famille G90/X6100 : 0x70.", 'cat.optYaesu': 'Yaesu (CAT natif)', 'cat.optKenwood': 'Kenwood (natif)', 'cat.civTrace': 'Journaliser le protocole CAT', 'cat.civTraceHint': '\u00c9crit dans le journal chaque trame CAT \u00e9chang\u00e9e avec la radio \u2014 CI-V en hexad\u00e9cimal, Kenwood en texte. Pour signaler une radio qui r\u00e9pond de travers \u2014 un bouton qui fait autre chose, une fr\u00e9quence qui saute. Valable pour la session seulement : c\u2019est un diagnostic, et le journal grossit vite.', 'cat.kenwoodPort': 'Port COM Kenwood', 'cat.kenwoodPortHint': 'Le port s\u00e9rie CAT/USB de la radio (TS-590, TS-890, TS-990, TS-2000, ainsi que les Elecraft K3/K4 qui parlent le m\u00eame dialecte).', 'cat.kenwoodBaudHint': 'Doit correspondre au MENU de la radio : un TS-590 sort d\u2019usine \u00e0 9600, un TS-890 \u00e0 115200.', 'cat.yaesuPort': 'Port CAT Yaesu', 'cat.yaesuPortHint': "Le port CAT de la radio — sur un FTDX10/FTDX101 en USB c'est le port COM ENHANCED, pas le standard.", 'cat.yaesuBaudHint': 'Doit correspondre au menu de la radio (FTDX10/FTDX101 par défaut : 38400).', 'cat.optIcom': 'Icom CI-V (USB série)', 'cat.optIcomNet': 'Icom CI-V (réseau / remote)', 'cat.optTci': 'TCI (Expert Electronics / SunSDR)',
   'cat.icomNetHost': 'IP / nom d\'hôte du poste', 'cat.icomNetUser': 'Utilisateur réseau (ID)', 'cat.icomNetPass': 'Mot de passe réseau',
   'cat.icomNetHint': "Se connecte directement au serveur LAN intégré du poste — sans RS-BA1 ni Remote Utility (ferme-les d'abord). Utilise l'ID/mot de passe Network User1 configurés dans le menu Network du poste. Un poste en veille est allumé automatiquement.",
   'cat.icomNetAudio': 'Diffuser l’audio RX par le réseau (expérimental)',
diff --git a/internal/cat/kenwood.go b/internal/cat/kenwood.go
index 0d73813..1c3d936 100644
--- a/internal/cat/kenwood.go
+++ b/internal/cat/kenwood.go
@@ -160,6 +160,34 @@ func (k *Kenwood) ReadState() (RigState, error) {
 	// IF reports the frequency of the VFO in USE (what the operator hears).
 	rx := f.FreqHz
 	tx := rx
+
+	// IF's split bit is not filled in by every rig that speaks this dialect —
+	// reported on a Flex through its Kenwood CAT emulation, where the frequency
+	// reads perfectly and split never appears. So ask the question directly as
+	// well: split IS "the transmit VFO differs from the receive VFO", which is
+	// what FR/FT answer, and it is the same rule the Yaesu backend settled on.
+	//
+	// A rig that rejects FR/FT answers "?;" once and is never asked again, so
+	// this costs two short commands per poll only where it actually works.
+	split := f.Split
+	if !split {
+		rxv, rxOK := k.askVFO("FR;")
+		txv, txOK := k.askVFO("FT;")
+		if rxOK && txOK && rxv != txv {
+			split = true
+			// Trust FR over IF for which VFO is in use: they were asked in the
+			// same breath, and a rig that leaves the split bit empty may be just
+			// as vague about the VFO field.
+			// f.VFO too, not just the reported state: the block below picks the
+			// TRANSMIT VFO as "the other one" from f.VFO, and leaving the two
+			// disagreeing would read the transmit frequency off the wrong dial.
+			if rxv == "A" || rxv == "B" {
+				k.curVFO, s.Vfo, f.VFO = rxv, rxv, rxv
+			}
+		}
+	}
+	f.Split = split
+
 	if f.Split {
 		// The transmit VFO is the other one. Read it rather than assume, and fall
 		// back to simplex if it cannot be read: a wrong TX frequency is written
@@ -235,6 +263,7 @@ func (k *Kenwood) write(cmd string) error {
 	if k.port == nil {
 		return fmt.Errorf("kenwood: not connected")
 	}
+	traceText("kenwood", "TX", cmd)
 	_, err := k.port.Write([]byte(cmd))
 	return err
 }
@@ -270,6 +299,7 @@ func (k *Kenwood) ask(cmd string) (string, error) {
 			}
 			frame := string(buf[:i+1])
 			buf = buf[i+1:]
+			traceText("kenwood", "RX", frame)
 			if frame == "?;" {
 				// The rig rejected the command. Remember it so the poll loop stops
 				// paying a 600 ms timeout for it on every cycle.
@@ -425,3 +455,30 @@ func (k *Kenwood) openPort() (serial.Port, error) {
 	}
 	return serial.Open(k.portName, &serial.Mode{BaudRate: k.baud})
 }
+
+// askVFO asks FR; or FT; and returns "A" or "B".
+//
+// The reply is FR0; / FR1; — the digit right after the two-letter command.
+// Anything else (a rig that answers with more fields, or not at all) returns
+// false, and the caller keeps whatever IF said rather than inventing a split.
+func (k *Kenwood) askVFO(cmd string) (string, bool) {
+	r, err := k.ask(cmd)
+	if err != nil {
+		return "", false
+	}
+	want := cmdPrefix(cmd)
+	body := strings.TrimSuffix(strings.TrimPrefix(r, want), ";")
+	if body == "" {
+		return "", false
+	}
+	switch body[0] {
+	case '0':
+		return "A", true
+	case '1':
+		return "B", true
+	}
+	// 2 is "sub receiver" on a TS-2000 — real, but not a VFO we track. Saying
+	// nothing is better than mapping it onto A or B and reporting a split that
+	// does not exist.
+	return "", false
+}
diff --git a/internal/cat/kenwood_loopback_test.go b/internal/cat/kenwood_loopback_test.go
index 2f500a2..133d8d9 100644
--- a/internal/cat/kenwood_loopback_test.go
+++ b/internal/cat/kenwood_loopback_test.go
@@ -108,7 +108,13 @@ type ts2000 struct {
 	mode       byte
 	onB        bool
 	split      bool
-	seen       []string
+	// lazyIF models a rig that answers FR/FT correctly but never fills IF's
+	// split bit — the behaviour reported on a Flex through its Kenwood CAT
+	// emulation, where the frequency reads perfectly and split never appears.
+	lazyIF bool
+	// noVFOCmds models a rig that rejects FR/FT outright ("?;").
+	noVFOCmds bool
+	seen      []string
 }
 
 func (r *ts2000) answer(cmd string) string {
@@ -127,7 +133,7 @@ func (r *ts2000) answer(cmd string) string {
 		return fmt.Sprintf("FB%011d;", r.vfoB)
 	case cmd == "IF;":
 		split := byte('0')
-		if r.split {
+		if r.split && !r.lazyIF {
 			split = '1'
 		}
 		// The catemu layout: IF | freq(11) | step(4) | RIT(±5) | 3 | mem(2) |
@@ -143,6 +149,21 @@ func (r *ts2000) answer(cmd string) string {
 	case strings.HasPrefix(cmd, "MD") && len(cmd) == 4:
 		r.mode = cmd[2]
 		return ""
+	case cmd == "FR;" || cmd == "FT;":
+		if r.noVFOCmds {
+			return "?;" // rejected, as a rig that does not know the command answers
+		}
+		// FR is the receive VFO, FT the transmit one. They differ exactly when
+		// the rig is in split.
+		v := vfoDigit
+		if cmd == "FT;" && r.split {
+			if vfoDigit == '0' {
+				v = '1'
+			} else {
+				v = '0'
+			}
+		}
+		return fmt.Sprintf("%s%c;", strings.TrimSuffix(cmd, ";"), v)
 	}
 	return "" // AI0;, TX;, RX; — set commands, no reply, as on a real rig
 }
@@ -251,3 +272,86 @@ func TestKenwoodSilentRigIsNotConnected(t *testing.T) {
 		t.Errorf("a stale model survived a failed connect: %q", k.model)
 	}
 }
+
+// Split found through FR/FT when the rig never fills IF's split bit.
+//
+// Reported on a Flex through its Kenwood CAT emulation: frequency read
+// perfectly, split never appeared. Rather than guess at which IF column that
+// firmware populates, ask the question that DEFINES split — is the transmit VFO
+// a different VFO from the receive one — which is what FR and FT answer, and
+// the same rule the Yaesu backend settled on after several wrong turns.
+func TestKenwoodSplitFromFRFT(t *testing.T) {
+	rig := &ts2000{vfoA: 14250000, vfoB: 14260000, mode: '2', lazyIF: true}
+	k := NewKenwood("COM-TEST", 9600, "FT8")
+	k.dialPort = dialTo(rig)
+	if err := k.Connect(); err != nil {
+		t.Fatalf("connect: %v", err)
+	}
+	defer k.Disconnect()
+
+	// Simplex: FR and FT agree, and nothing may be invented from that.
+	s, err := k.ReadState()
+	if err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	if s.Split {
+		t.Errorf("split reported while FR and FT agree: tx=%d rx=%d", s.FreqHz, s.RxFreqHz)
+	}
+
+	// Split on, IF still silent about it: receive on A, transmit on B.
+	rig.split = true
+	if s, err = k.ReadState(); err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	if !s.Split {
+		t.Fatal("split not detected — FR/FT disagreed and IF's bit was empty, which is the reported case")
+	}
+	if s.FreqHz != 14260000 || s.RxFreqHz != 14250000 {
+		t.Errorf("tx=%d rx=%d — want tx 14260000 (B), rx 14250000 (A)", s.FreqHz, s.RxFreqHz)
+	}
+}
+
+// A rig that rejects FR/FT is asked once, then left alone — and its IF split
+// bit still works. The fallback must not cost a timeout on every poll, nor
+// break the rigs that were already fine.
+func TestKenwoodSplitWhenFRFTRejected(t *testing.T) {
+	rig := &ts2000{vfoA: 14250000, vfoB: 14260000, mode: '2', noVFOCmds: true}
+	k := NewKenwood("COM-TEST", 9600, "FT8")
+	k.dialPort = dialTo(rig)
+	if err := k.Connect(); err != nil {
+		t.Fatalf("connect: %v", err)
+	}
+	defer k.Disconnect()
+
+	if _, err := k.ReadState(); err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	asked := 0
+	for _, c := range rig.seen {
+		if c == "FR;" || c == "FT;" {
+			asked++
+		}
+	}
+	if _, err := k.ReadState(); err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	after := 0
+	for _, c := range rig.seen {
+		if c == "FR;" || c == "FT;" {
+			after++
+		}
+	}
+	if after != asked {
+		t.Errorf("a rejected command was asked again: %d then %d", asked, after)
+	}
+
+	// IF's own split bit still drives the result on such a rig.
+	rig.split = true
+	s, err := k.ReadState()
+	if err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	if !s.Split || s.FreqHz != 14260000 || s.RxFreqHz != 14250000 {
+		t.Errorf("IF-reported split broke: split=%v tx=%d rx=%d", s.Split, s.FreqHz, s.RxFreqHz)
+	}
+}
diff --git a/internal/cat/log.go b/internal/cat/log.go
index f195b47..5a4359b 100644
--- a/internal/cat/log.go
+++ b/internal/cat/log.go
@@ -75,37 +75,53 @@ func DebugLogPath() string {
 	return filepath.Join(base, "OpsLog", "cat.log")
 }
 
-// ── CI-V byte trace ────────────────────────────────────────────────────────
+// ── CAT wire trace ─────────────────────────────────────────
 //
-// Opt-in, off by default. Turning it on logs every CI-V frame sent and received
-// as hex, exactly like the WinKeyer protocol trace.
+// Opt-in, off by default. Turning it on logs every frame sent and received —
+// CI-V as hex, the ASCII backends (Kenwood, Yaesu) as the text they exchange.
 //
-// That trace exists because a fault nobody could reason about — "the keyer sends
-// one element then stalls" — was settled in one line the moment the actual bytes
-// were visible. The CI-V link has now produced the same class of report: a MOX
-// button that sets split instead of transmitting, on a rig whose PTT command is
-// demonstrably correct in the source. Guessing at that from the command table
-// has already cost this project a wrong fix; the bytes will say what the rig was
-// really asked.
-var civTrace atomic.Bool
+// It exists because a fault nobody could reason about — "the keyer sends one
+// element then stalls" — was settled in one line the moment the actual bytes
+// were visible, and the CAT links have since produced the same class of report:
+// a MOX button that sets split instead of transmitting, and a split the Kenwood
+// backend does not recognise on a rig whose frequency it reads perfectly.
+//
+// Both are questions about what the RIG actually said. Guessing at that from a
+// protocol document has already cost this project a wrong fix that broke the
+// case which already worked.
+var catTrace atomic.Bool
 
-// SetCIVTrace turns the CI-V byte trace on or off.
+// SetCIVTrace turns the CAT wire trace on or off. Named for the CI-V link it
+// was written for; it now covers the text backends too.
 func SetCIVTrace(on bool) {
-	civTrace.Store(on)
+	catTrace.Store(on)
 	if on {
-		debugLog.Printf("civ trace ON — every CI-V frame to and from the rig is logged")
+		debugLog.Printf("wire trace ON — every CAT frame to and from the rig is logged")
 	} else {
-		debugLog.Printf("civ trace OFF")
+		debugLog.Printf("wire trace OFF")
 	}
 }
 
 // CIVTraceEnabled reports whether the trace is running (for the settings UI).
-func CIVTraceEnabled() bool { return civTrace.Load() }
+func CIVTraceEnabled() bool { return catTrace.Load() }
 
-// traceCIV logs one frame. dir is "TX" (to the rig) or "RX" (from it).
+// traceCIV logs one CI-V frame. dir is "TX" (to the rig) or "RX" (from it).
 func traceCIV(dir string, b []byte) {
-	if !civTrace.Load() || len(b) == 0 {
+	if !catTrace.Load() || len(b) == 0 {
 		return
 	}
 	debugLog.Printf("civ %s % X", dir, b)
 }
+
+// traceText logs one ASCII exchange, for the Kenwood and Yaesu backends.
+//
+// The text is QUOTED. A truncated or empty reply must be visible as such
+// rather than vanish into the line: "" and ";" look identical unquoted, and
+// telling them apart is the whole question when a rig answers a command it
+// does not really support.
+func traceText(backend, dir, s string) {
+	if !catTrace.Load() || s == "" {
+		return
+	}
+	debugLog.Printf("%s %s %q", backend, dir, s)
+}