The slope released into the Bhote Koshi headwaters at 08:37. Two seismometers registered it within 43 seconds — one 57 km away, one 131 km away — while the water took another 18 minutes to reach the first settlement. Thirteen more stations, out to 260 km, recorded it well enough to time, and together they place the source within 1.3 km. The waveforms identify the event as a mass movement and show the slope releasing a second time three hours later. Seismic monitoring is the one channel that saw this disaster while it was still upstream.
ms_vx, depth 0 km, reviewedTwo instruments carry the record. NK.KKN at Kakani is a three-component Guralp CMG-3T sampling at 50 Hz, 57 km from the source. IO.EVN at the Pyramid Laboratory in the Khumbu is a CMG-3ESPC at 20 Hz, 131 km out on a different bearing. The traces below are one-minute RMS ground velocity with the instrument response removed, so the two are directly comparable. Both peak within a minute of each other at 08:37–08:38, and the signal stays above the quiet-period range for 285 minutes. At rest Kakani sits at 21 nm/s and Khumbu at 38 nm/s.
One-minute RMS ground velocity, 1–8 Hz, instrument response removed. Log scale, 25–26 Aug 2026, Nepal time.
The onset is emergent and the disturbance persists for hours — the signature of mass
moving down a valley. A tectonic source produces an impulsive arrival that decays within
minutes. USGS typed the event landslide on the ms_vx scale,
which is reserved for mass-wasting sources.
| Time (NPT) | Duration | NK.KKN | IO.EVN | What happened |
|---|---|---|---|---|
| 08:12–08:27 | 15 min | 178 nm/s | 348 nm/s | Elevated signal before the catalogued origin. Could be progressive failure, could be unrelated noise. |
| 08:28 | — | 31 nm/s | 89 nm/s | Continuous elevated ground motion begins. |
| 08:37 | peak | 6.80 µm/s | 13.28 µm/s | USGS origin 08:37:10, M5.2 ms_vx, type=landslide, depth 0 km, reviewed. |
| 11:45 | secondary | 89 nm/s | 451 nm/s | Second catalogued failure, M4.2, same location — the slope released again while the first signal was still elevated. |
| 13:13 | 285 min total | — | — | Signal returns to the station baseline. |
Ground motion rises from 08:12 to 08:27, ending ten minutes before the catalogued origin time. Progressive failure of the slope before release would produce this, and so would rainfall or an unrelated local source. Separating those requires testing the same signature against the second failure and against the wider archive.
The seismic location is 28.2710 N, 85.5150 E — inside the Timure catchment, 14 km from the settlement. Checked against our own terrain and glacier layers, that point sits at 5,590 m with 3,706 m of relief within 3.5 km, on slopes averaging 41°, with an RGI glacier 40 m away and three glaciers within 2 km. That is ice–rock avalanche terrain, and it matches the seismic signature.
Every point on the blue curve is 73 km further from Khumbu than from Kakani, which is what a 24.3 s delay means at a 3.0 km/s crustal group velocity. The dashed pair either side spans 2.8–3.2 km/s. Hover anywhere to see the delay a source at that spot would have produced.
Basemap: Esri. Distances use a local equirectangular approximation, accurate to better than 1 km at this scale.
Against the water's own clock, from the moment the slope let go.
Seismic travel times at 3.0 km/s; flood-front arrival from the along-channel distance at 15–20 m/s. Time axis is compressed so seconds and minutes both stay readable.
One delay gives a curve, not a point. The source lies somewhere on it. The branch does run through the glaciated headwaters above the Bhote Koshi, and the amplitude ratio independently puts the source nearer Kakani, which together narrow it to a short arc of that curve.
As an after-the-fact check only: the location USGS published hours later sits 1.3 km from this curve. That number plays no part in drawing it.
A survey of every open station within 82–90 E found 19 with data over the event window and 15 that recorded it well enough to time — one Guralp broadband, one CMG-3ESPC, and thirteen Raspberry Shake units between 49 and 260 km. That is enough to stop guessing at a curve and invert for a position.
Amplitude cannot be used for this. Site response varies by a factor of eleven between stations with effectively the same noise floor — Khumbu amplifies this event 3.0× while a Shake at Gorkha damps it to 0.27× — and one event cannot separate fifteen unknown site terms from the source. Arrival-time differences are indifferent to how much a site amplifies, so the inversion uses those: each station's 1–8 Hz envelope cross-correlated against Kakani, then a search over position and velocity with origin time divided out.
| Station | km to source | Envelope lag vs NK.KKN | cc |
|---|---|---|---|
| AM.R3B1E | 50 | -2.0 s | 0.90 |
| NK.KKN | 57 | reference | — |
| AM.RF015 | 73 | +5.0 s | 0.97 |
| AM.R2C6D | 77 | +6.0 s | 0.86 |
| AM.RE2F2 | 127 | +7.0 s | 0.76 |
| AM.RA5AB | 93 | +8.0 s | 0.90 |
| AM.S8618 | 111 | +12.0 s | 0.94 |
| AM.R02C2 | 75 | +14.0 s | 0.86 |
| AM.RC951 | 184 | +15.0 s | 0.88 |
| IO.EVN | 132 | +22.0 s | 0.95 |
| AM.R5A68 | 136 | +23.0 s | 0.90 |
| AM.RB628 | 209 | +28.0 s | 0.86 |
| AM.R0F23 | 144 | +29.0 s | 0.91 |
| AM.R43A3 | 237 | +49.0 s | 0.91 |
| AM.RCCCC | 261 | +60.0 s | 0.91 |
Every station correlates at 0.76 or better. The recovered Kakani–Khumbu difference is +22.0 s, against the 24.3 s measured directly from the two waveforms above — independent agreement from a different method.
The orange region is where the arrival-time differences alone allow the source to be. It is long north–south and narrow east–west because every station lies south of the source, so azimuth is well determined and range is not. The blue patches are the only ground inside it that could produce an ice–rock avalanche — slope 30° or more, above 4,000 m, within 2 km of a glacier. Click any station or patch for its numbers.
Region contours are RMS-misfit screening levels, not formal confidence intervals. Terrain from Copernicus GLO-30 and RGI 7.0. Basemap: Esri.
Timing alone puts the source at 28.370 N, 85.420 E — 14.4 km from the epicentre USGS later published, which falls inside the region. The region's shape is the network's signature: 79 km north–south against 39 km east–west, because all fifteen stations sit in a 170° arc to the south, west and east. The remaining 190° is Tibet and will never hold a station, so range will always be the poorly determined direction here.
Terrain fixes exactly that. Of the 2,432 km² the timing allows, 497 km² is steep enough, high enough and close enough to ice to fail as an ice–rock avalanche — and it breaks into 972 disconnected patches. Ranked by area, the largest is the right one: 72 km² centred 1.3 km from the USGS epicentre, peaking at 7,205 m with a mean slope of 47°. It is also the closest of the top forty, against a median of 27 km. A disc that close covers 0.22% of the search region, so landing there by chance is roughly one in 450.
The contours are RMS-misfit screening levels, not formal confidence intervals, and the best-fit velocity of 3.80 km/s sits high for Rayleigh waves — emergent onsets and one-second envelope sampling both blunt the timing. The terrain thresholds were set to match this scar as already described, so scoring this same event with them is mildly circular. Above all it is one event: the method recovered a known answer once. That is a demonstration, not a measured hit rate. The M4.2 three hours later is the test that would make it one.
A second failure, M4.2, was catalogued from the same location at 11:45:35 — three hours after the first, while ground motion from the first was still elevated. The source slope did not stabilise after one release.
The seismic wave reached Kakani 9 seconds after the failure. The water took 18 minutes to reach the nearest settlement.
Along-channel distance from the source at 5,590 m, with the flood-wave arrival window at 15–20 m/s. The seismic wave covered the 57 km to Kakani in about 9 seconds.
Timure would have had 18–24 minutes; Betrawati 51–68 minutes. Turning that into warning would have required a process watching the stream continuously and a route to reach people downstream in under twenty minutes. Neither was in place on the morning of the 26th, and neither is in place now.
The record runs continuously to 31 August at both stations. No further failure of comparable size appears, and the USGS catalogue adds nothing at this location after 11:45 on the 26th. Daily medians sit within the pre-event range on the 27th and 28th.
Ten-minute RMS ground velocity, 1–8 Hz, response removed. Both panels share one time axis and one scale. Day rules are Nepal midnights; breaks in a line are telemetry gaps.
One episode does stand out. From 02:45 on 30 August Kakani rises to 188 nm/s, nine times its resting level, holds through 04:45, and decays smoothly until about 10:45 that morning, 8 hours in all. The onset is gradual, with no sharp arrival. It is below the catalogue's threshold and carries no USGS entry.
Two explanations were tested and neither holds. Rainfall over the corridor correlates at r = −0.05 across the period, and the amplitude climbs while rain is easing off. Cultural noise is ruled out by the clock: that same pre-dawn window on the other days in the record is the quietest of the day, running 14–17 nm/s at Kakani, against 88 nm/s through this episode, 5.5 times the quiet-day level for the same hours.
IO.EVN, the Pyramid Laboratory in the Khumbu, sits 131 km from the source on a different bearing. It peaks at 13.3 µm/s during the avalanche against Kakani's 6.8 µm/s, and reaches 129 nm/s during the 30 August episode against Kakani's 188 nm/s.
Khumbu reading higher during the avalanche is a site effect, not proximity. Across three quiet days its amplitude runs a steady 1.95× Kakani's, and the two resting levels — 38 versus 21 nm/s — carry the same ratio. That factor is a property of the site and the instrument, so it has to be divided out before the two numbers say anything about distance.
Once it is, the ratio becomes informative. The avalanche gives Khumbu/Kakani of 2.9, well above the 1.95 site factor. The 30 August episode gives 0.69, far below it — Kakani reads more than Khumbu in absolute terms, which for two stations with this site ratio means a source substantially nearer Kakani. That is the Rasuwa side.
From 20:05 on 30 August Khumbu climbs, peaking at 231 nm/s — 6.1× its resting level — while Kakani holds at 16 nm/s, its own resting level. A signal at one station with nothing at the other is local to that station, so this one belongs to the Khumbu rather than the Rasuwa corridor. Read from Khumbu alone it would look like a second event.
Extending the record settles it: the excursion decayed and was back inside its quiet band by 02:45 on 31 August, 6.7 hours in all, and Khumbu has stayed there for the 19 hours since. It was an episode, not the onset of something.
What has continued is different in kind. The background at Khumbu is back to quiet — median 44 nm/s across the 112 bins since the episode closed, against a resting level of 38 — but discrete transients keep arriving on top of it. 12 of those 112 bins exceed twice the quiet level, 11%, against 0.7% on 28 August, a quiet reference day. The largest reaches 440 nm/s at 19:35 on 31 August — bigger than anything inside the episode itself.
Most are not artifacts. Checked individually against the raw waveform, ten of the twelve — including the largest — sit in continuous data with no telemetry break within twenty minutes either side, so they are not filter ringing at a segment edge. The remaining two have a dropout nearby and are set aside. Kakani records no bin above twice its own quiet level on the 31st, so they are local to Khumbu. Short, discrete, repeating events beside a station at 5,010 m are what rockfall looks like. The sustained episode ended; whatever is producing these has not.
Corroboration is not location. A later survey of the region found thirteen further stations carrying data, so the network is not as thin as two — but none of them helps here. The only one near Khumbu, a Raspberry Shake 31.6 km away, has no data between 14:00 and 17:30 UTC on the 30th, which is the episode itself; it was not a silent witness but an absent one. The genuinely quiet instruments in the network sit 205 and 387 km from Khumbu, far too distant for a source this small. Two amplitude measurements on one flank give two intersecting distance circles and no way to choose between them, and the episode's peak arrives about an hour apart at the two stations, which fits a sustained process more than a single impulsive source.