On the morning of 26 August 2026 a mountainside above the Lhende Khola, in Rasuwa, fell into the gorge of a small river. Seven minutes and forty seconds later a front of ice, rock and mud, moving at about 170 km/h, destroyed the border crossing at Rasuwagadhi, 22 km downstream, met the main river there, and ran the length of the Bhote Koshi and the Trishuli as a flood. I ask one narrow question with physics: where did all that water come from? It is answered below, and the answer is an energy budget rather than a model. A second question — how big the collapse was — is not answered. Seven versions of the routing model have failed the tests I set them, and on 8 September I stopped building versions. Both states are on this page.
As of 8 September, 1,342 people are confirmed dead in Nepal and 4,886 missing, with 43 dead and 519 missing on the Chinese side, and more than 5,300 injured. The missing count has begun to fall while the death toll rises — that is bodies being recovered, not people being found alive. Of roughly 900 recovered by 1 September, about 4% had been identified. Many of the dead were hydropower workers, from the plant at the border to the tunnels at Upper Trishuli-1. Every figure on this site is provisional.
The wave carried away bridges, vehicles, homes and people. No term in this arithmetic stands for them, and none of the numbers on these pages should be read as if it did.
दिवंगतहरूको स्मृतिमा — in memory of those lost.
What this is, before you read any of it. Independent, AI-assisted analysis, published days after the event by one person working outside his own field. I am an engineer in New Zealand's electricity sector, not a geoscientist. It is not peer-reviewed. No Nepali scientist has read it. By its own licence it is not fit to sit under an operational warning system, an evacuation decision or a hazard zoning without independent expert review. Five results have been withdrawn so far and all five are still readable. Everything here is preliminary, and the numbers that depend on the routing model are more preliminary than the rest. The contribution is arithmetic on figures already public — the relative sizes of the terms — not a finding about how well anyone else has done their job.
Every finding carries a confidence chip, and the colour is the part that means something. green the conclusion survives every combination of inputs I can still defend. blue it has survived the tests run against it, including a named challenge, but it rests on at least one input that could still move. amber unresolved, under revision, or partly withdrawn. The chip says how hard a finding has been shot at, not how much I believe it.
The water was already in the riverholds across the range
Melting ice is expensive and falling is cheap. A kilogram
falling 1,200 m releases about a 28th of the heat needed to melt a
kilogram of ice. Frictional melt therefore caps out at
1.1 Mm³ on best evidence, or 2.5 with
every dial turned in melt's favour, against the ~20 Mm³ that Nepal's
Flood Forecasting Division measured above base flow at Devghat between
14:10 and 18:00 on the day. Something had to supply that, and melting ice
cannot. It was the monsoon-swollen river itself, swept up and delivered at
once — from the border junction down. Above the junction the Lhende held
about 0.5 Mm³ of water, so the first 22 km were the mountain, not
a flood.
Caveats. This is an energy budget and it does not depend on the
routing model, which is the reason it has not moved while everything else
has. It does depend on the collapse size in finding 04, which is under
revision and probably too small — but a larger collapse raises the melt
ceiling in proportion, and the conclusion survives until the size reaches
roughly 10× its current top. Melt only reaches 20 Mm³ at
200 Mm³ of collapse, 80% ice, the full 4,000 m drop and a heat
partition at the very top of the published range, all at once.
It did not fall like dry rockdry rock excluded
The border camera's clock — 7 min 40 s for
22 km — turns out to measure what the falling mass was made of,
not how big it was. Dry-rock versions arrive at 17–30 minutes whatever
size I make them; ice-rich ones arrive within minutes of the observed time.
The arithmetic works without any model: the drop over the run is
H/L = 0.14, so it only arrives at all if its effective friction is
below 0.14, and Scheidegger's rule gives dry rock 0.28–0.30 at this size,
which stops it after about 10 km. This ran 22.
Caveats. The clock rests on a CCTV timestamp and a seismic origin
time, both of which I have had to reconcile, and an earlier version of this
page overstated how independently it was corroborated. One published speed
(19 m/s, peer-reviewed) contradicts the ones I scored against; I think
it measures the water after the turn rather than the front, but that
reconciliation is mine and unreviewed, so both scorings are published.
The method travels — one test stands, one is withdrawnone of two stands
Against Chamoli 2021 the model reproduces the arrival time
at Tapovan once frictional melting is given a thermal lag, with one fitted
number and the rest held. Against Seti 2012 it does not: that test was
withdrawn on 3 September after I found a bad channel profile underneath it —
straight lines flown across a mountain rim, then flattened to zero slope by
the routine meant to clean them. Rerun on a corrected channel the test
fails and its second conclusion reverses.
Caveats. Nothing about the Trishuli findings depends on Seti. The
Chamoli test shares a channel-profile pipeline with this analysis, so it is
not fully independent of it.
How big the collapse was — unresolved, and may be withdrawnunder revision
This finding used to say the collapse was far smaller than
the headline figures, at 14–34 Mm³. I no longer believe
that. Three things since say it is too small: the volume that ran
3 km up the Chinese side and came back, a mud line 45–70 m above
the bed at Upper Trishuli-1, and the mud-line map itself, which needs a
release nearer 110–175 Mm³ in this model. Two
independent lines published since would make that ordinary rather than
large: a source area of 1.009 km² implies a slab 109–173 m thick,
and Chamoli was about 135; and another group's estimate of ~100 Mm³
±40% overlaps it from below.
But I have not replaced the number. Seven versions of the routing
model have been scored against the mapped water depths and none has
survived. On 8 September I stopped building versions, because a check
of my own scoring showed the single run that ever passed all 11 tests
matches the mud lines at 11% of the points in the first 22 km — it hit
every reach average while being wrong at most places inside them. That is a
question not asked precisely enough to have an answer, not a search that
needs one more attempt. This finding may end up withdrawn rather
than corrected. Everything on these pages that depends on it should
be read as conditional; finding 01 does not.
I proved my own model wrong, then fixed itfalsified, then fixed
The model could drop sediment but never pick any up.
Adding erosion — with formulas published for other rivers on other
continents and nothing tuned to this event — gave 3.8 Mm³ of scour
against 3.2 measured by satellite stereo. It also broke the model: with
deposition working properly the flood stalled before reaching the towns,
and 150 runs across every plausible combination could not both arrive on
time and leave the right debris. The fix was to stop treating sediment as
one substance — boulders settle and grind, silt and ice ride with the water
and leave no measurable deposit.
Caveats. An independent analyst, Darcy Weedman at geopera, reached
the same conclusion the same week by a different route. I lean on geopera's
satellite work heavily — the erosion figure, the border superelevation
speed, the deposit location — so "a different method" is true of the solver
and not of the inputs.
What is not settledopen
What the mass was made of is not resolved: wetness and ice content span nearly the whole range I allowed, and an earlier version of this page appeared to settle them, which turned out to be an artefact of holding one dial fixed. The modelled deposit sits at km 0–36 where stereo measurement puts it at 40–43. The height the river rose at Galchhi fails out of sample. The Devghat peak passes on a factor-of-2 test, but every surviving run lands below the observation, which is a one-sided residual rather than a clean pass. And no Nepali scientist has read any of this.
My contribution is quantification, not correction. The scientists working on this event — Kargel, Shugar, Petley, ICIMOD and others — have consistently described river water and entrained sediment in their accounts of the flood. What nobody had published was the arithmetic: the relative sizes of the terms. That is the gap this project set out to fill, and where my numbers disagree with an agency's, both are shown. Nothing here is a finding about how well anyone else has done their job, and the warning-system section of the plain-English page is about a gap in the design of national systems generally, not about the people who were operating one that morning while four of its gauges were being destroyed.
8 September. The seventh version of the
routing model failed, and I stopped building versions. The reason was a check
on my own scoring: the one run that ever passed all 11 tests matches the mapped
mud lines at 11% of the points in the first 22 km. It satisfied every reach
average while being wrong at most places inside them. The work is now split in
two at the junction — an avalanche above, a flood wave below — so each half can
be tested on its own. Also on 8 September: the first academic reconstruction of
the collapse appeared and agrees with my geometry, and I corrected a figure I
had quoted with a precision its source never claimed.
Earlier: 5 findings withdrawn, several corrected, 7 model
versions built and none promoted.
The full record, with dates and reasons →
The plan changed on the evening of 8 September and this section changed with it. For a week the answer here was "another version of the model". Seven have now been built, none has survived its tests, and the reason turned out to lie in how the model was being scored as much as in the model (changelog above). The work is now split in two. The first twenty-two kilometres — a rock-and-ice avalanche in a confined gorge, tested against the border camera's clock and the speed read off its frames — are one problem. The river below the junction, which gathers its water as it travels rather than carrying it, is another. They meet at a single handover: the shape of the flood as it passes the junction. The first thing to test needs no avalanche model at all — whether any shape of flood at the junction can be routed down the river the mud lines describe. Either answer is worth having, and a "no" would say something about my reconstruction, or about the lower river's physics, that no further version could. Before any of that, two smaller jobs: a filtering bug that has been quietly discarding about a fifth of the mud-line measurements, and scoring the model against the whole measured profile instead of six reach averages. Results go to the changelog first, as the rule above requires.
Not a request — a statement of what the work does not have, so that a reader can weigh it. The DHM gauge records at Betrawati (station 447) and Galchhi both survived the flood; a peak discharge at either would settle the one question this model cannot, which is how much of the flood that passed Betrawati reached Galchhi and how fast. I do not have them, and I have not asked for them. The same goes for the hydropower headworks and SCADA records across the thirteen affected projects, which plausibly hold the only high-frequency measurements of the upper reach where the public gauges were destroyed; for clear-sky stereo or drone photogrammetry of the glacier below the scar, which would test directly where the falling mass got its mobility and which nobody has published; and for India's CWC hourly archive on the Gandak, which would extend the routing past the border. Everything on this site is built from what is already public, and its limits are the limits of that.
No Nepali scientist has reviewed this. It is the most conspicuous gap in the work. The Nepali translation is withdrawn rather than published on machine translation.