diff --git a/app/changelog.json b/app/changelog.json
index 4dd15b10..f839c43e 100644
--- a/app/changelog.json
+++ b/app/changelog.json
@@ -6,6 +6,13 @@
     "title": "Node positions straight after a server restart",
     "body": "The server fetches the node registry on its own, so nobody sees it as unreachable after a restart."
   },
+  {
+    "id": "2026-09-30-quieter-send-sounds",
+    "date": "2026-09-30",
+    "where": "app",
+    "title": "Quieter send sounds",
+    "body": "The pops for Discover and trace-pings are a third as loud and shorter, so a weak reception is heard over them."
+  },
   {
     "id": "2026-09-27-hud-bottom-space",
     "date": "2026-09-27",
diff --git a/app/src/__tests__/soundengine.test.js b/app/src/__tests__/soundengine.test.js
index 3a34fa13..19969762 100644
--- a/app/src/__tests__/soundengine.test.js
+++ b/app/src/__tests__/soundengine.test.js
@@ -1,5 +1,5 @@
 import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest'
-import { createSoundEngine } from '../sound.js'
+import { createSoundEngine, cuePeak, TX_POP } from '../sound.js'
 
 // A plain zero-hop cue: these tests are about the engine's clock and
 // lifecycle, not about which instrument plays (#468).
@@ -56,7 +56,7 @@ function makeCtx({ state = 'running', sampleRate = 48000 } = {}) {
     close() {},
     createGain: () => node({ gain: fakeParam(1) }),
     createBiquadFilter: () => node({ type: '', frequency: fakeParam(0), Q: fakeParam(0) }),
-    createConvolver: () => { ctx.convolvers++; return node({ buffer: null }) },
+    createConvolver: () => { ctx.convolvers++; return node({ buffer: null, isConvolver: true }) },
     createStereoPanner: () => { ctx.panners++; return node({ pan: fakeParam(0) }) },
     createBuffer: (ch, len) => ({ getChannelData: () => new Float32Array(len) }),
     createBufferSource: () => { ctx.bufferSources++; const n = node({ buffer: null, loop: false, start() {}, stop() {} }); ctx.sources.push(n); return n },
@@ -588,3 +588,56 @@ describe('the ambient layer has depth, and gives it all back', () => {
   })
 
 })
+
+// #602: with sound on, auto-discover was the loudest thing in the mix. A sweep
+// of N targets is N + 2 pops every ten seconds, each at 0.16 through the full
+// reverb, over receptions whose dit scales with RSSI and sits well under that
+// when weak. The cue for "a frame went out" carries no measurement; the cue
+// for "something was heard" does, so that is the one that has to be on top.
+describe('a transmit pop sits under the receptions (#602)', () => {
+  // How much of a source reaches the reverb, and how much reaches the output
+  // without it: the product of the gains along every path, read off the fake's
+  // wiring. An envelope gain counts as 1, its resting value in the fake.
+  function shares(from) {
+    let wet = 0, dry = 0
+    const walk = (n, level, viaReverb) => {
+      if (n === ctx.destination) { if (viaReverb) wet += level; else dry += level; return }
+      const here = n.gain ? level * n.gain.value : level
+      for (const next of n.outs) walk(next, here, viaReverb || !!n.isConvolver)
+    }
+    walk(from, 1, false)
+    return { wet, dry }
+  }
+  const lastOsc = () => ctx.oscillators[ctx.oscillators.length - 1]
+
+  it('peaks below the quietest direct dit of a weak reception', () => {
+    // The issue's own measure: at -110 dBm the network voice, the driest and
+    // most common one, must come out over the pop.
+    expect(cuePeak('network', -110)).toBeGreaterThan(TX_POP.peak)
+    for (const family of ['advert', 'channel', 'message', 'trace']) {
+      expect(cuePeak(family, -110), family).toBeGreaterThan(TX_POP.peak)
+    }
+  })
+
+  it('is what cue() plays a dit at, so the comparison above is about the engine', () => {
+    const e = createSoundEngine()
+    e.setMode('rxtx')
+    const ramps = []
+    const gain = ctx.createGain
+    ctx.createGain = () => { const g = gain(); g.gain.linearRampToValueAtTime = (v) => { ramps.push(v); return g.gain }; return g }
+    e.cue(DIRECT, -110)
+    expect(ramps).toContain(cuePeak('network', -110))
+  })
+
+  it('sends a fraction of itself to the reverb where a dit sends all of it, at the same dry level', () => {
+    const e = createSoundEngine()
+    e.setMode('rxtx')
+    e.cue(DIRECT, -80)
+    const dit = shares(lastOsc())
+    e.txBlip('trace')
+    const pop = shares(lastOsc())
+    expect(dit.wet).toBeGreaterThan(0)
+    expect(pop.wet / dit.wet).toBeCloseTo(TX_POP.reverb, 5)
+    expect(pop.dry).toBeCloseTo(dit.dry, 5)
+  })
+})
diff --git a/app/src/sound.js b/app/src/sound.js
index b218fc12..e61daf01 100644
--- a/app/src/sound.js
+++ b/app/src/sound.js
@@ -198,6 +198,26 @@ const MUSIC_DENSITY = 1.5 // periods divided by this — how often notes fall
 const RX_GAIN = 0.5       // reception dits, independent of the music/bed level
 const FADE_S = 0.03       // music-bus fade before cutting voices, avoids a click
 
+// cuePeak is the level a direct dit is played at: the RSSI's loudness, the rx
+// share of the mix and the family's own gain. Exported because the transmit
+// pop is set against it (TX_POP): what is on top in the mix is a relation
+// between the two, not a property of either.
+export function cuePeak(family, rssi, offset = 0) {
+  return pingGain(rssi, offset) * RX_GAIN * (VOICES[family] || VOICES.network).gain
+}
+
+// The transmit pop (#602), chosen by ear in a lab on the real engine, Kasper
+// 2026-09-30. It was 0.16 with a 160 ms tail through the whole reverb, and a
+// sweep of N targets is N + 2 of them every ten seconds: the cue that carries
+// no measurement was the loudest and most frequent thing in the mix. At 0.05 a
+// direct network dit at -110 dBm (0.078) comes out 1.6 times over it, where
+// the pop used to be 2.1 times over the dit.
+//   peak    gain at the top of the pop's envelope
+//   tail    seconds from the start to silence
+//   reverb  the share of the pop that goes through the master, and so into its
+//           reverb send; the rest goes out dry at the same level
+export const TX_POP = { peak: 0.05, tail: 0.07, reverb: 0.15 }
+
 // Generative music (Eno's Music-for-Airports technique): seven pad voices,
 // each looping ONE note of a calm F-pentatonic set on a mutually prime period.
 // The periods share no common divisor, so the combination never repeats.
@@ -243,6 +263,7 @@ export function createSoundEngine({ random = Math.random } = {}) {
 
   let ctx = null, mode = 'off', bed = null, cueState = {}
   let master = null, genTimers = [], genGain = null, activeOscs = []
+  let txDry = null
 
   // Created lazily from the FAB tap (a user gesture, which Web Audio requires).
   // If the context comes back suspended anyway (persisted mode restored at boot,
@@ -264,6 +285,12 @@ export function createSoundEngine({ random = Math.random } = {}) {
       out.gain.value = 0.9
       lp.connect(out).connect(ctx.destination)
       master = lp
+      // The master's rounding without its reverb send, for the share of a
+      // transmit pop that stays dry (TX_POP.reverb).
+      txDry = ctx.createBiquadFilter()
+      txDry.type = 'lowpass'
+      txDry.frequency.value = 6500
+      txDry.connect(out)
       // Reverb: synthesized impulse response (decaying noise), constant wet
       // send — part of the approved sound, not a runtime setting.
       const len = Math.floor(ctx.sampleRate * REVERB_SECONDS)
@@ -608,7 +635,7 @@ export function createSoundEngine({ random = Math.random } = {}) {
     const v = VOICES[c.family] || VOICES.network
     const f = harmFreq(rssi, offset)
     const frac = rssiFrac(rssi, offset)
-    const g = pingGain(rssi, offset) * RX_GAIN * v.gain * (c.damped ? DAMP.gain : 1)
+    const g = cuePeak(c.family, rssi, offset) * (c.damped ? DAMP.gain : 1)
     const len = (0.035 + frac * 0.06) * v.hold * (c.damped ? DAMP.hold : 1)
     const t = ac.currentTime
 
@@ -708,11 +735,17 @@ export function createSoundEngine({ random = Math.random } = {}) {
     osc.frequency.setValueAtTime(f * 0.55, when)
     osc.frequency.exponentialRampToValueAtTime(f, when + 0.05)
     gain.gain.setValueAtTime(0, when)
-    gain.gain.linearRampToValueAtTime(0.16, when + 0.01)
-    gain.gain.exponentialRampToValueAtTime(0.0001, when + 0.16)
-    osc.connect(gain).connect(master)
+    gain.gain.linearRampToValueAtTime(TX_POP.peak, when + 0.01)
+    gain.gain.exponentialRampToValueAtTime(0.0001, when + TX_POP.tail)
+    osc.connect(gain)
+    const wet = c.createGain()
+    wet.gain.value = TX_POP.reverb
+    gain.connect(wet).connect(master)
+    const dry = c.createGain()
+    dry.gain.value = 1 - TX_POP.reverb
+    gain.connect(dry).connect(txDry)
     osc.start(when)
-    osc.stop(when + 0.2)
+    osc.stop(when + TX_POP.tail + 0.04)
   }
 
   function txBlip(kind) {
@@ -759,7 +792,7 @@ export function createSoundEngine({ random = Math.random } = {}) {
   function destroy() {
     stopBed()
     stopMusic()
-    if (ctx) { try { ctx.close() } catch (_) {} ctx = null; master = null; genGain = null }
+    if (ctx) { try { ctx.close() } catch (_) {} ctx = null; master = null; txDry = null; genGain = null }
   }
 
   return { setMode, cue, txBlip, destroy }
diff --git a/changelog.d/2026-09-30-02-quieter-send-sounds.json b/changelog.d/2026-09-30-02-quieter-send-sounds.json
new file mode 100644
index 00000000..71ba6cb1
--- /dev/null
+++ b/changelog.d/2026-09-30-02-quieter-send-sounds.json
@@ -0,0 +1,7 @@
+{
+  "id": "2026-09-30-quieter-send-sounds",
+  "date": "2026-09-30",
+  "where": "app",
+  "title": "Quieter send sounds",
+  "body": "The pops for Discover and trace-pings are a third as loud and shorter, so a weak reception is heard over them."
+}
diff --git a/docs/2026-07-16-sound-modes.md b/docs/2026-07-16-sound-modes.md
index 57396e66..cfd26639 100644
--- a/docs/2026-07-16-sound-modes.md
+++ b/docs/2026-07-16-sound-modes.md
@@ -63,6 +63,16 @@ singing-bowl and sampled-ambiance beds; real music tracks vs. generative.
   density, reverb **wet 35% / decay 2.8 s** on pings, tx cues, and music.
 - Mix constants live at the top of `sound.js` (`REVERB_*`, `MUSIC_*`, `RX_GAIN`).
 
+> **Amended 2026-09-30 (#602): the transmit pop sits under the receptions.** With sound on, auto-discover was the loudest thing in the mix: a sweep of N targets is N + 2 pops every ten seconds, each at gain 0.16 with a 160 ms tail through the whole reverb, while a dit scales with RSSI and a weak one sits well under that. The cue for "a frame went out" carries no measurement; the cue for "something was heard" does. Chosen by ear in a lab on the real engine (Kasper, 2026-09-30), `TX_POP` in `sound.js`:
+>
+> | | was | is |
+> |---|---|---|
+> | peak gain | 0.16 | 0.05 |
+> | tail | 160 ms | 70 ms |
+> | share through the master and its reverb send | 100% | 15%, the rest dry at the same level |
+>
+> Measured at the master for a direct network dit at -110 dBm (0.078, `cuePeak`): the pop was 2.1 times over it, and the dit is now 1.6 times over the pop. Every transmission still sounds: the Discover broadcast as two pops, each trace-ping as one. Sounding only the Discover pops, and a silent auto-discover, were both in the lab and not chosen. The rule of #254 is untouched: the cue follows the frame, not the timer.
+
 ## Behaviour choices
 
 - Pings follow the **filtered/plotted set** plus `hops === 0` — you hear what
diff --git a/web/changelog.json b/web/changelog.json
index 4dd15b10..f839c43e 100644
--- a/web/changelog.json
+++ b/web/changelog.json
@@ -6,6 +6,13 @@
     "title": "Node positions straight after a server restart",
     "body": "The server fetches the node registry on its own, so nobody sees it as unreachable after a restart."
   },
+  {
+    "id": "2026-09-30-quieter-send-sounds",
+    "date": "2026-09-30",
+    "where": "app",
+    "title": "Quieter send sounds",
+    "body": "The pops for Discover and trace-pings are a third as loud and shorter, so a weak reception is heard over them."
+  },
   {
     "id": "2026-09-27-hud-bottom-space",
     "date": "2026-09-27",