The Trishuli Water Ledger

00Status — read this first

First published 2 September 2026. Sections 01–14 are substantially as written then; where a later result has overtaken one, it is flagged in place rather than silently rewritten. The full running record — every result withdrawn or corrected, with the date and the reason — is on the changelog page, and the working record is PLAN.md in the repository. It is not repeated here.

00bThe event, as far as it is known

A description, not an argument. Every number is sourced in research/event-dossier.md; where sources disagree the range is given.

The mountain. Langtang Lirung is a 7,227 m peak about 60 km north of Kathmandu, in Rasuwa. Its north face stands above the Chhochen Khola, a headwater of the Lhende Khola, about 4 km inside Nepal. At 08:37:10 on 26 August 2026 — clear, rain-free, late monsoon — a section of that face roughly 600 m wide and 0.2 km² let go from about 5,200 m. It registered as a magnitude-5 seismic event; the Department of Mines and Geology recorded M4.4 at 08:37 near Rasuwagadhi, and M2.6 at 08:20 — a 17-minute precursor. Chinese researchers have separately reported unusual signals at ~06:04, ~06:50 and 08:15–08:19. Every version of the routing model assumes a single release; a staged failure is untested.

What fell. Bedrock failure that took a hanging glacier with it — the consensus of everyone who has examined the imagery. The ice fraction is unpublished. Volume estimates span a factor of 400. The first academic reconstruction maps the source area at 1.009 km² and declines to give a volume for want of a co-registered elevation pair (§04b).

The path. ~1,200 m of drop onto moraine and loose ground, scoured on the way, then another ~1,000 m to the stream bed, then 22 km down the Lhende in 7 min 40 s — a mean 47.8 m/s, and it did not decelerate. For the last 14 km that gorge is the Nepal–China border. It destroyed the Rasuwagadhi crossing and the hydropower plant there, met the Bhote Koshi, ran ~3 km up the Kyirong arm into Tibet and back, and turned down the Bhote Koshi and the Trishuli as a flood.

The gauge record (Flood Forecasting Division, 27 August release; transcribed in full in research/ffd-press-release-27aug.md):

Warning. The Division issued its first warning at 09:15 and sent 679,295 SMS alerts to riverbank residents — 38 minutes after the collapse, by which time the wave had passed everything above Syabrubesi, and between two and six hours ahead of the river crossing its warning levels at Malekhu, Kalikhola and Devghat. In the first 22 km no chain that passes through a person could have been quick enough at 7 min 40.

Deposits and marks. Satellite stereo puts the main deposit at km 40–43 where the valley opens, with 11–18 m of bed rise at the Upper Trishuli-1 headworks. Mud lines mapped along both banks on 7 September give a peak-stage profile for the whole corridor (output/trimlines.csv) — 73 m [59–99] in the gorge, ~68–90 at Hakubesi, falling to single figures below Galchhi. A debris dam in the upper valley held a lake for two days before draining.

01The question

When part of Langtang Lirung's north face let go — bedrock plus a hanging glacier; published guesses now span ~1–10 million m³ for the initial detachment to an unsourced 100–200 million m³ total, and nobody has yet published a firm volume or ice fraction — the mass fell about 1,200 m, hit the valley floor, and arrived at the China–Nepal border 22 km away seven minutes and forty seconds later (six minutes fifty until 6 September — see §00). What followed down the Bhote Koshi and Trishuli was not a landslide but a flood: a sediment-charged wave that gauges tracked all the way to the Narayani confluence.

The question this page tries to answer with arithmetic rather than adjectives: where did that much water come from? The narrative that leads most coverage — inherited from the well-studied 2021 Chamoli disaster — is that the fall's energy frictionally melted the glacier ice. My hypothesis, formed watching the videos: the debris wave acted as a moving dam, sweeping up the monsoon river that was already in the channel, kilometre after kilometre, and that cumulative river water dominates the ledger.

To be fair to the scientists quoted in the press: Jeff Kargel's own account names "melted avalanche ice, entrained sediment and the river itself", and Dave Petley's is similar. Nobody serious is ignoring river water. But nobody has published the relative fractions either — and the fractions are the whole story. Dan Shugar put it plainly: "Figuring out where that water came from is actually probably the hardest thing about all this."

02The corridor

Geography first, because it is routinely reported wrongly. The collapse was in Nepal — the scar sits about 4 km inside the international boundary, on the north face of Langtang Lirung, and fell into the Chhochen Khola, a headwater of the Lhende. But from roughly km 8 the Lhende is the boundary: over the next 14 km the modelled path never departs from the mapped border by more than 180 m, Nepal on one bank and Tibet on the other. The Lhende then dead-ends into the main Kyirong Tsangpo–Trishuli stem — a river that has itself come out of Tibet — at the crossing where Gyirong Port sat directly opposite the mouth, on the far bank from Rasuwagadhi. So the flood was transboundary within eight minutes, and a substantial share of the water it swept up had entered the system on the Chinese side of the boundary — which matters for the ledger below, because that water is part of the standing channel volume the wave collected. From there the river is Nepal's Bhote Koshi, then the Trishuli, gathering tributaries until it meets the Kali Gandaki at Devghat to form the Narayani. The junctions on this map are more than scenery: when the surge arrives, an entire side arm's storage suddenly connects and the flow pays a momentum toll to turn — a piecewise step in the system, of which the border T-junction is the first and sharpest (§08).

03The clocks

The event left a chain of timestamps — a CCTV camera, four drowned gauging stations, one surviving gauge, and an official peak — now anchored by the Department of Hydrology and Meteorology / Flood Forecasting Division (DHM/FFD) press release of 27 August (a scanned two-page PDF obtained via the Wayback Machine). They are the calibration targets for everything below. Distances are measured along the river channel I stitched from OpenStreetMap (199 km scar→Devghat; the FFD quotes 168 km on a shorter reckoning).

One end of that first interval has been audited and the other has not. On 6 September I stopped trusting a third party's account of the Gyirong overlay and read the footage frame by frame, which is what moved the arrival from 08:44:00 to 08:44:50. The start — 08:37:10 — is still the USGS catalogue origin for us7000tbwb, taken on trust. That event was reclassified from an M4.4 tectonic earthquake to an M5.2 landslide-generated signal, and a landslide radiates an emergent onset rather than the impulsive first arrival a catalogue origin-time solution is built around; no force-time inversion has been published that would pin it independently. The uncertainty there is seconds, not minutes, so no conclusion in this report turns on it — but the correction made earlier the same day was itself only 50 seconds and moved the headline speed by 12%, which is the reason to name this rather than leave it implied. Every "minutes after the collapse" on this site inherits it.

Reading the camera: the clock, the calibration, and a direct speed

Added 6 September. Method and measurement, logged in the changelog rather than promoted to a finding. Full working: research/event-dossier.md §14; tooling in research/video/. The footage is third-party and not in the repository.

The footage. A 1280×720 upload of the Gyirong Port CCTV, 66.9 s long, carrying the station's own burnt-in clock. Its frame rate was the first thing I got wrong: a perceptual duplicate-frame filter reported 23.85 unique frames per second, and a strict pixel-identity test gave 48–57 — 57 of 60 during the arrival. One frame is 16.7 ms, not the 33 ms an earlier 30 fps resample had assumed, and that resample was itself throwing away half the information. Everything below uses the original frames at their original timestamps.

The clock. The overlay ticks in whole seconds, but the tick edges are detectable and fall at video t = n + 0.100 s; the gaps between detected digit changes were 3.033, 0.967 and 2.000 s, all within 35 ms of exact integers, which is what shows the clip is real-time and uncut. That pins video t = 0 to 10:59:25.900 Beijing, ±35 ms. First dust at the left of frame at 10:59:47.5; the black plume emerges from behind the building at video 24.700 s — 10:59:50.6 Beijing, 08:44:50.6 Nepal — which is the arrival used in the table above. One caveat the footage cannot close: the CCTV clock's own drift is unknown.

The calibration, two ways. The border-post facade is mirror-symmetric about the national emblem, and the dentil frieze under its roofline is a graduated scale: Dave Hume counted 18 squares from the centre square to the building's end, over a 33.5 m half-length, giving a 1.97 m module (a 2 m architectural set-out; automated peak detection found the same 18, with the centre square within one pixel of the emblem). A projective fit to those graduations had an rms of 1.97 px. Independently, he surveyed the two building ends from the camera position in Google Earth: bearing 167.82° at 117 m and 194.00° at 145 m.

Two unrelated routes agreeing to 2–3% is what makes the rest usable. The derived camera has a focal length of about 1,411 px, a 48.8° horizontal field of view, ~14° of downward tilt and an optical axis on bearing 192.22°, which reproduces the surveyed 194.00° at the right-hand building edge exactly. The local image scale runs from 11.05 px/m at the left of frame to 7.62 at the right — a 45% gradient — so no single pixels-per-metre factor is valid across this picture, and every conversion goes through the projective map.

The measurement. The debris front crosses sight line 1 at source frame 1482 (24.700 s) and sight line 2 at about frame 1546 (25.766 s). The second crossing is ambiguous between two adjacent frames — the front is hard to separate from a shadow in the cliff behind — and the later, slower reading was taken deliberately; no sub-frame interpolation was done. That choice is worth 2.3 m/s. Fifty metres in 1.066 s is 46.9 m/s — 169 km/h.

Timing contributes about 1.6%; the separation is everything else, and its range is asymmetric (−5/+15 m) because 50 is the best reading of where the sight lines touch the ground rather than the midpoint of what is defensible. Quote it as 41–63 m/s, nominal 47. It is an instantaneous front speed, one sample of a quantity that fluctuates as obstacles fail. One item stays open: the range at which the 50 m was read. If it was taken at the building (~145 m) while the debris was passing at ~200 m, the true separation is nearer 69 m and the speed nearer 68 m/s — the sight lines diverge, so the gap grows with range.

What it agrees with, and what it excludes.

The envelope contains every independent estimate of the front and excludes the one estimate of the post-turn water — which is the discrimination that matters for §08, and it survives the full distance uncertainty. A caution about the top three: 46.9, 47.8 and 48.5 agreeing within 1.6 m/s by three unrelated routes is tighter than a ±10 m distance uncertainty can support. Read it as consistency, not precision; part of the three-way agreement is luck. The run-up geometry does not discriminate inside the envelope either — 46.9 m/s needs a stagnation efficiency of 0.45 to produce the observed 45–55 m of cliff run-up, and across 41–63 that efficiency runs 0.22–0.63, physical throughout — but the CAS 19 would need 2.7, which is impossible, and that is the point: it is not the same quantity. Note also that the instantaneous ~47 at km 22 equals the 47.8 mean over the whole 22 km: the front did not slow down the Lhende. The collapse to ~21 m/s over km 22–68 happened at and after the junction.

What failed, so that it is not repeated. Tracking the black plume laterally does not work: six frieze graduations spanning 106 px all darkened within a quarter of a frame, and the sequence runs backwards at the end — the cloud appears, billowing toward the camera, it does not translate, and a robust line fit through those crossings returned 66–87 m/s on data that violates its model. The white plume and the black plume are different phenomena; every edge-tracking number produced before that was noticed was measuring the white one. Edge tracking across every threshold, window and set-back spanned 12–58 m/s, which contains every published estimate and is therefore not a measurement. And a wedge-geometry reconstruction with the river only ~23° off the line of sight amplified every pixel of error into metres; it returned 95–170 m of travel for a plume crossing a 65 m building, a red flag rather than a result. Three claims made during that work were withdrawn the same day and are listed in the dossier.

04The energy wedge

Melting ice is expensive: 334 kJ per kilogram. A kilogram of anything falling 1,200 m releases just 11.8 kJ. So even if every joule of the initial fall became heat and every joule of that heat went into ice, the collapse could melt no more than ~3.5% of its own mass. Frictional melt is energy-limited, not ice-limited — it doesn't matter how much ice fell; it matters how far it fell.

I tested this bookkeeping against the one well-instrumented analogue. For Chamoli 2021 (26.9 Mm³, 80:20 rock:ice, 3,400 m fall — winter, rivers near empty), my budget reproduces the Science paper's published answer: ~5 Mm³ of meltwater, ice-limited, with river water a ~8% term. Shugar, Kargel and colleagues called that rock:ice ratio "almost exactly the critical value required for near-complete melting" — a rare coincidence, not a general rule.

Run the same bookkeeping on Langtang 2026, and the answer depends entirely on how big the collapse was — which is the one thing nobody has measured, so I run it across the range. The two highlighted rows are where my own evidence puts it; the last row is the case the melt narrative needs:

Drop in the table is the vertical fall charged to melt energy: 2,400 m is scar to the Lhende channel — the ~1,200 m initial fall plus the ~1,000 m below the impact zone (§00b); 4,000 m is scar to the lower river, the most generous case. The "1,200 m" quoted on the summary pages is the initial fall only.

So melt reaches the FFD's ~20 Mm³ of excess at exactly one point in the whole parameter space: the largest published volume — six times the top of my own envelope — with four-fifths ice, the longest drop and a heat-to-ice partition at the top of the literature's 0.3–0.7 range, all at once. Relax any single one and it falls away fast. At the size my own consistency envelope supports, melt is 1.1 Mm³, or 2.5 Mm³ steel-manned: short of the official excess by a factor of eight. Reproduce with calcs/energy_water_budget.py, scenarios 5–8.

Correction, 5 September. Until today this section led with the 100 Mm³ row and the summary pages said the ceiling sat "far below" 20 while §13 conceded 15–20 — a contradiction produced by leaving the size envelope in the finding that generated it instead of carrying it here. The conclusion is unchanged; the margin is roughly eight times larger than I was claiming, and now rests on a stated case rather than an adjective.

Meanwhile the monsoon corridor held roughly 22.4 Mm³ of water standing in the channel (300–600 m³/s flowing at ~3 m/s over 168 km, the FFD's reckoning of the distance), delivered ~10 Mm³ more of baseflow during the seven hours the wave took to pass, and its saturated bed sediments carried ~9 Mm³ of pore water. The wedge between those numbers is the whole argument. And the early-September estimates trend smaller still: if the initial detachment was ~5–10 Mm³ (the EGU hydrology blog's order-of-magnitude; Jakob Steiner's elevation-data reading, which also raises the initial fall to ~2,100 m), the melt term shrinks below 1 Mm³ and the wave's growth becomes almost entirely an entrainment story — the snowplow in its purest form.

04bIce and rock together are not rock: the first 22 kilometres

Added 6 September. The first 22 km were not a flood. The Lhende is small: about 150 m³/s where it meets the Bhote Koshi, less higher up, and all the water standing in its channel that morning comes to about 0.5 Mm³. Add pore water from the gravel it tore up and the little friction had melted, and 2–4 Mm³ is the ceiling — against a release of tens of millions. What struck the border was at least four-fifths mountain. That reframes finding 02 without moving it and sharpens where finding 01 applies: below the junction, where 177 km of river held some 22 Mm³ of water. "Wall of water" was the wrong description of what arrived at km 22.

So what got a mountain 22 km in 7 min 40 s?

Ice — not as water, as ice — and the simplest way to see it uses no model. For a mass sliding from the scar to the junction floor, ~3,100 m of descent over 22 km, the drop-to-run ratio is H/L = 0.14. A sliding block arrives at all only if its effective friction is below that, and arrives still moving at 47 m/s only if μ is very close to it, about 0.135. Scheidegger's volume–mobility regression, the standard rule for dry rock avalanches, gives μ ≈ 0.30 at 21 Mm³ and 0.28 at 34: dry rock this size stops after about 10 km. That is finding 02 restated without a routing model, and it is why the dry-rock scenarios arrive at 17–30 minutes only by borrowing channel water on the way down.

The literature is more careful than the slogan. Rotating-drum experiments (Schneider, Kaitna et al. 2011) found granular ice cuts a gravel mass's bulk friction by up to 20 % and water by about 50 %, with mixtures above ~40 % ice nearly liquefying — but in a field set of 64 rock–ice avalanches over 1 Mm³ the same group could not confirm ice fraction as a control on runout; what showed up was water, volume, and travel over a glacier. Rock avalanches running on glaciers give Voellmy friction of 0.03–0.10 (Sosio et al. 2012) and ice-rich glacier detachments reach ~5°, H/L ≈ 0.09 (Kääb et al. 2021). The model's ice-rich scenarios use μ = 0.17, attributed in the code to Schneider et al.; I could not open that paper's full text and say so.

The closest analogue is Kolka–Karmadon, Caucasus, 20 September 2002: 100–130 Mm³, roughly 80–90 % glacier ice, which ran 18–20 km at a mean 50–80 m/s with a drop-to-run ratio of 0.15, stalled where the valley geometry stopped it, and released a mudflow that ran on another 15 km. The Lhende run is that event at a fraction of the size — similar distance, similar mean speed, similar ratio, a debris body that stopped where the geometry stopped it, and a flood that continued. Huascarán 1970 is the other classic (50–80 Mm³, 13–16 km in 3–4½ minutes, mobility attributed to entrained till, snow and meltwater). Chamoli 2021 is the contrast: at 80 % rock, a similar ratio but a front already down to ~25 m/s by 15 km and 34–37 minutes to Tapovan — what a mostly-rock mass with a thin ice budget does. One further piece of Chamoli arithmetic carries over: a 3,100 m drop can melt at most about a fifth of a mass's volume if that fifth is ice, so anything ice-richer arrives at the bottom still carrying ice — which is what the border footage shows.

Why ice does this — the physics of an ice skate

Not, mostly, pressure melting: that lowers the melting point by 0.0074 °C per atmosphere, and skating works at −30 °C. Three mechanisms, in the order that matters at avalanche speeds. Premelting — an ice surface carries a disordered, liquid-like film with no load at all (Faraday; Dash, Rempel & Wettlaufer 2006 is the modern review), thickening toward the melting point. Frictional melting — the power dissipated per square metre of sliding contact is large enough at these speeds to melt a continuous film, whatever the bulk temperature. And fragmentation — the ice breaks up as it travels, so the mass keeps generating fresh surfaces rather than polishing one.

An independent measurement of the same ratio — added 8 September

A preprint posted this month (arXiv 2609.04563) reconstructs this collapse from open data and reports a mean endpoint angle of 8.84° — the same quantity as H/L: tan 8.84° = 0.156. Its geometry is independent of mine: its crown, toe, path length and elevation model. My own §15 arithmetic from a 5,200 m crown gives 0.154, or 8.76° — about 1 % apart. (The 0.14 above is measured to the junction floor rather than the debris toe; both are on the page because they answer different questions.)

What that is worth. Every speed here is a distance over a time and every slope a drop over a distance, all measured along my centreline on my chainage, so a stretched path would corrupt the lot quietly. This is the first time an outside group has measured the same quantity and landed on my answer, and it says the geometry is sound. But it is a simple measurement — similar public elevation data, a crown and toe that are fairly obvious on an image — so agreement is close to what one should expect. It validates the foundation, not the building. It says nothing about whether the volume is right, nothing about the velocities, and nothing about the routing model.

Their source area turns the volume into a thickness, and the thickness is ordinary. The same preprint maps the detachment at 1.009 km² (0.491–1.841 across its delineations) and declines to publish a volume, on the stated grounds that no suitable co-registered pre- and post-event elevation pair exists.

So the number that reads as uncomfortably large is an entirely ordinary slab for this class of failure, and that arrives from a source which took care not to estimate a volume. Two further numbers point the same way: a boulder census in the deposit reach (4,329 clasts over a metre, largest resolved 26 m) gives a competence velocity of 19–26 m/s, inside the window this model is scored against where the depth measurement is strongest; and geopera's released volume of ~100 Mm³ ±40 % overlaps my range from below. The preprint's twelve corridor transects — 145 to 1,298 m wide, median 533 — are the flooded valley floor of §12b measured by a third method.

On the range of published volumes. Estimates in circulation run from geopera's ~100 Mm³ ±40 % to my 110–175, and figures near 200 Mm³ are also reported. On the arXiv source area of 1.009 km², 200 Mm³ is a slab about 196 m thick — still ordinary for this class of failure, so nothing in the physics rules it out. One caution about the range 197–492 Mm³ as it appears in general reference sources: the arithmetic resembles a detachment area multiplied by a drop height rather than a detached volume (0.2 km² × 1.2 km = 240 Mm³). I have not been able to trace that particular range to a named study, which is a limit of my searching and not a judgement on anyone's figure. I do not use it, and I do not describe my own number as low relative to it.

Where this leaves the two pieces of work. The preprint declines a dynamic model explicitly, on the grounds that one "would add false precision without source volume, material fractions, hydrograph or suitable validation". I built one anyway and constrained the hydrograph from mud lines, and it has failed the tests I set it, in public, seven times — which is some evidence for their caution. Other groups are modelling this event dynamically too; an earlier version of this paragraph implied an empty field, which was wrong.

05The snowplow model

To go beyond static budgets I built a one-dimensional routing model along the real channel: 499 elevation samples down the stitched OSM river path (Copernicus- derived terrain via the Mapzen dataset). The front advances at a speed set by local slope plus a flood-wave floor, U = 300·S^0.82 + 4 m/s, calibrated against the clocks above. As it advances it integrates the river: standing channel water is swept into the wave, baseflow keeps arriving, pore water is liberated from the bed the flow tears up, and the meltwater slug from the steep upper reach rides along. The wave body travels slower than the front (fitted factor 0.90), so the pulse stretches — about 45 minutes long at the border, ~3.3 hours long at Devghat, matching the FFD's observed 15:20 arrival → 16:00 peak → 18:30 return to normal — and the peak decays except where re-fed.

06What the ledger says

Four scenarios, one test: reproduce the only two hard downstream numbers — the 5,850 m³/s peak at Devghat and the Flood Forecasting Division's ~20 Mm³ of "excess" water.

Pulse shape (duration growth and the fraction of active water riding in the main window) is calibrated once against the FFD's Devghat stage record and shared across all four scenarios — so the discriminator is the ratio between rows, not the absolute tuning. Melt-only cannot reach the observed peak even with the same generous shape. Late caveat (2 Sept): the FFD's volume/duration arithmetic suggests the Devghat gauge base flow is ~2,900 m³/s (Narayani — see §07), not the ~1,500 assumed in this table's peak column; on that reading the observed excess is ~2,900 m³/s, the snowplow overshoots it and the ladder undershoots it, and the verdicts will be re-scored once the gauge and rating are confirmed (§13). The volume-based discrimination — melt-only cannot supply the FFD's 20 Mm³ — is unaffected.

Melt-only physics cannot make this flood. Even granting melt three times the energy budget's best estimate, the wave arrives at Devghat a fifth too small. The scenarios that work are the ones where close to four-fifths of the water in the distal wave is channel-derived — swept-up monsoon flow plus collected baseflow — with saturated-sediment pore water second and glacier melt a ~8% term in that v1 run (at the 14–34 Mm³ envelope it is a few per cent). In the upper gorge, where the world's cameras pointed, there was almost no river water to displace (§04b); with every kilometre downstream the river's share compounds.

07Gross versus net — the subtlety in the 20 Mm³

One bookkeeping distinction does a lot of work. Water the wave merely redistributes — channel water concentrated from 199 km of river into a three-hour pulse — shows up at a gauge as a huge surge, but over a day-scale window it nets toward zero: the emptied channel refills from the baseflow behind it. Genuinely new water — melt, and pore water squeezed from entrained sediment — does not cancel. My snowplow scenario puts the gross surge at ~65 Mm³ and net-new water at ~10 Mm³.

The FFD's press release resolves what their "20 million cubic metres of excess" means: it is the flow above base level measured at Devghat between 14:10 and 18:00 — a windowed gross excess at one station, marked preliminary. That definition captures redistributed channel water that arrived inside the window and misses the long tail. Run their arithmetic backwards and it also reveals the gauge: 20 Mm³ over the ~3.8-hour window is a mean excess of ~1,450 m³/s — a triangular peak excess of ~2,900 — implying a base flow near 2,950 m³/s at the gauge. That is the Narayani below the Kali Gandaki confluence (Trishuli ~1,500 plus Kali Gandaki ~1,450), not the Trishuli arm alone — a reading their numbers make mutually consistent, and one that changes what "peak 5,850" demands of a routing model (§08, §13). Either way, the gap between the gross wave and the net-new water is itself the evidence: redistribution, not creation, is what made this flood enormous. A melt-only event has no such gap.

What this figure can and cannot test. It is not a volume balance on the collapse, and it must not be set against a release volume: a windowed gross excess at the bottom gauge counts redistributed channel water that happened to arrive inside the window, misses the tail that arrived after 18:00, and says nothing about what was deposited, ponded or still frozen upstream. What it is is a constraint on the shape of the wave at Devghat, and that is something the ensemble has never been scored on. Every version so far has been tested on the height of the Devghat peak and never on its width; ~20 Mm³ above base between 14:10 and 18:00 fixes the area under the excess hydrograph over a stated window. Wave width is precisely what a flood spreading across a valley floor changes (§12b), so this goes into the next version as a test — scored loosely, at one significant figure, as the source states it.

08The river as a circuit

There is a second way to see all of this, natural to anyone who works on power systems. Drop the inertia term from the shallow-water equations and what remains — the diffusive wave approximation used in flood routing since the 1930s — is mathematically identical to a nonlinear RC transmission line: the telegrapher's equations without the inductor. Stage is voltage, discharge is current, Manning friction is a nonlinear series resistor, and valley storage is shunt capacitance — tiny in the gorges, huge where valleys open. The observed collapse in speed from ~50 to ~11 m/s at Syabrubesi is a fat shunt capacitor clipping the pulse.

Side valleys are RC branches: the passing surge drives water up each tributary mouth, charging, and it drains back after the peak — clipping the crest and feeding the tail. A temporary debris dam is a breakdown element, a capacitor behind a rising resistance that fails at threshold like an SCR firing; a chain of them is a relaxation-oscillator cascade, which is what the 2012 Seti flood's ~27 successive surges look like.

The analogy is not decoration. It is why the model is built as a node network with an explicit junction element, and it is what makes the junction step at km 22 — where the Lhende meets the Bhote Koshi — a modelled component rather than a boundary condition.

09Downstream into India

The wave did not stop at Devghat. At the Valmikinagar barrage on the Gandak (the Narayani's Indian name), roughly 90 river-kilometres further on, Bihar's control-room figures — transcribed in local press — show the arrival: from a ~2,900–3,100 m³/s base, discharge climbed from 20:30 IST to a flat peak of 4,253 m³/s (150,200 cusecs) at 23:00–23:30 IST, the highest figure recorded there that day, then receded overnight. All 36 gates were open. That is a lag of ~7¼ hours from the Devghat peak and ~27% attenuation — and a celerity of only ~3.4 m/s, half the hill-reach value, exactly what the circuit picture predicts when the line runs onto the Chitwan floodplain: a wide, flat, fat-capacitance section. (An earlier spike in the barrage record at 16:00 IST is almost certainly a gate-operation transient, not routed flow — it is simultaneous with the Devghat peak 90 km upstream.)

The 28 August barrier-lake overtopping in the headwaters — feared as a second flood — is not resolvable at the barrage: fluctuations stayed within ~±150 m³/s of normal. Consistent with the ladder's lesson: by 300 km, the RC cascade has low-pass-filtered small pulses into the noise. The fine multi-pulse structure, if it existed, lived in the reaches where the gauges were destroyed; the Galchhi record and India's CWC hourly archive are where it would still be written.

10Does it travel? The Chamoli stress test

Every fitted law flatters the event it was fitted to. If this pipeline is to be worth anything as a warning tool it has to travel — so I took it, frozen, to the most hostile documented case: Chamoli, 7 February 2021. Winter, a near-dry channel, a 26.9 Mm³ mass that was ~80% rock — the opposite of a monsoon river valley in every regime variable — and ground truth of rare quality (a seismic detachment time, video-tracked speeds at three points, a 34–37 minute arrival at Tapovan 23 km downstream; Shugar et al., Science). The Trishuli-fitted kinematic law, applied verbatim, fails exactly as regime thinking predicts: it delivers the front to Tapovan in 11 minutes — three times too fast — while passing all three local speed checks. A slope-only law has no flow regime and no event size; Chamoli is precisely the event that exposes both omissions.

So I routed the actual release block down the actual profile with the ladder’s momentum equation (§08) plus one new term: Coulomb basal friction, which turns the friction law into Voellmy’s — and collapses back to Manning (and to every published chart in this report, bit for bit) when μ = 0. The dissection came out clean. A constant μ fails in both directions at once: a hundred-metre-deep flow outruns Coulomb friction in the steep upper canyon, while the flat basin at Raini (slope ~0.015) is terminal — the front stalls there and never reaches Tapovan. And the observations themselves contain the diagnosis: the mean speed to Tapovan (~11 m/s) is below the slowest of the three local speeds (12 m/s) — arithmetic that forces an intermittent, stop-and-go front, surging when moving and stopped in between.

What closes the gap is Shugar et al.’s own mechanism, made dynamic. The ~20% ice melted progressively under frictional heating — their energy budget (and mine, §04) shows the fall’s energy was almost exactly sufficient to melt all of it — so the friction dial μ slides from a dry-granular value toward a watery one as the mass descends. Both endpoints exist in the literature untouched by Chamoli: Scheidegger’s 1973 volume–mobility law gives μ ≈ 0.30 for a dry rock avalanche of this volume, and the wet endpoint (~0.02) is this report’s own Manning limit. One physical constant remains: dissipated heat takes minutes to conduct into ice fragments before it melts them, so the melt state rides with the mass and lags the fall. That thermal lag τ is the one degree of freedom I fit, and the observed arrival selects τ ≈ 5 minutes — ice debris of ~6 cm, a plausible comminuted size. The rest is then out of sample: arrival 36.3 min against an observed 34–37 — inside the window rather than beside it; mean speed 10.6 m/s against ~11; the km-16 flow speed passes. The basin crossing is where it now reads worst: 13 minutes against the ~27 that an independent analysis (Rana et al.) published for that reach. That gap widened after I fixed an interpolation bug in the arrival calculation and re-ran — the arrival itself improved, the basin transit got worse, and both are reported here rather than only the first. The model still cannot do stick-slip — it smears the surging into steady creep, so the two local speeds nearest Tapovan read low — and I say so rather than fit harder: that microstructure needs friction hysteresis or two-phase mechanics (Shugar’s team needed full r.avaflow to do better than this).

11The climate question

Much of the coverage frames this as a climate disaster. For this event, that framing claims more than the evidence holds — and so does its mirror image, "nothing to do with climate." The two claims need separating.

What the record does not support: this was not a weather event (a clear, rain-free morning), not a glacial-lake outburst, and — per the ledger above — not even mostly glacial water: melt was a ~8% term in the v1 ledger and, at the 14–34 Mm³ envelope of §00, at most 1.1–2.5 Mm³ — a few per cent. The proximate cause was a bedrock wedge failure of a kind steep, tectonically active Himalayan faces have produced throughout the Holocene; USGS notes an earlier event on the same massif in 2015 — that one off the south flank, 7 km away and into a different valley, so "the same mountain" rather than "the same location" (§02). And much of the death toll is an exposure story — a border post, a customs yard, and hundreds of hydropower workers from 13 projects packed into one corridor, in the reach that no warning could have reached in time. No formal attribution study exists yet; any causal percentage quoted this week is ahead of the science.

What is genuinely climate-linked — mechanism, not hand-waving: the failure zone sits at ~5,200 m, in the elevation band where warming permafrost degrades the ice that cements jointed rock — the best-documented driver of increasing high-mountain rock-slope failure. Locally, glacierised areas above 4,000 m are warming ~0.31 °C per decade; Langtang's glaciers have retreated 1.5× faster since 2000 than before; Lirung itself is back 450 m since 1990, debuttressing the slopes it once held up. The warmest ground and lake temperatures in two years were observed in the weeks before failure, with meltwater filling crevasses — the same water-at-the-bed mechanism blamed for the 2016 Aru glacier collapses in Tibet. And the neighbouring headwater produced a supraglacial-lake outburst thirteen months earlier — one of the most directly warming-linked hazard types in the Himalaya. Two disasters in one corridor in thirteen months is anecdote, not statistics; it is also what the trend literature predicts.

12For the hydropower cascade

This corridor carried one of the densest hydropower cascades in the Himalaya: 13 projects — 748 MW built or building — took damage, 431 MW came off Nepal's grid, and the missing include on the order of nine hundred hydropower workers, from the 111 MW Rasuwagadhi plant at the border T-junction to the hundreds caught in Upper Trishuli-1's tunnel (~350 of whom were rescued alive on 28 August). Every one of these projects is run-of-river — small weirs, hours of pondage at most. There was no storage reservoir anywhere on the system capable of absorbing a surge (the basin's one large storage proposal, the 1,200 MW Budhi Gandaki, is unbuilt): in circuit terms the cascade added no meaningful capacitance to the line — only exposure. For readers in the industry, three things in this analysis bear on how such a corridor behaves:

— Up-valley, the wave outruns any human alert chain. The front covered the first 22 km in seven minutes forty; the alert went out 38 minutes after the collapse, in time for the lower river and necessarily too late for the upper. In the upper reach the only thing that operates on that timescale is automation — a tripwire wired directly into plant control — because a human loop cannot close in seven minutes. Below Betrawati the wave took hours, and the alert chain did work there.
— Confluences are exposure hot-spots. The junction step (§08) concentrates impact, backwater and momentum exchange exactly where cascade infrastructure tends to sit — at tributary mouths. Rasuwagadhi's headworks stood 400 m below one.
— Plant records are the missing instrument. Headworks stage sensors, gate logs and SCADA historians across the thirteen projects — including drowned plants whose servers survived — plausibly hold the only high-frequency hydrographs of the event's upper reach, better than anything the washed-out public gauges left behind. Nothing in this report is built on them; that is a statement of what this analysis is missing, not a proposal to anyone.
— The trigger can extend weeks earlier. Sentinel-1 InSAR shows the failed mass creeping ~10 mm/month and accelerating in the days before collapse (Shirzaei, Virginia Tech, via Nature news, 2 Sept) — a satellite watchlist tier above the seismic tripwire. Note that a second, independent reading of the same satellite record (geopera, 31 Aug) finds under two metres of steady motion in a hundred days and no acceleration at all; the two disagree, neither is peer-reviewed, and precursory creep should not be treated as established. And GFZ's Niels Hovius has said seismic-signal-based downstream warning was feasible "with only minutes' delay" — the automated-tier case, made independently.

12bThe collapse was larger than the published envelope — and I cannot yet say by how much

Three routes, none sharing an observable with the 14–34 Mm³ envelope of §00, say that envelope is too small.

1. The volume that ran up the Chinese arm. Treat the junction as a node: measure what the flow parked up the Kyirong arm at its peak, estimate the share that went that way, divide, and subtract the Lhende's own contribution. Sixty-one cross-sections up the arm on the Copernicus 30 m DEM, filled to the mapped mud line, give 23 Mm³ (15–33) in the pond alone. The share is the crux and the geometry cannot pin it: anywhere from a tenth to a half. At a half the through-volume lands at the top of the envelope; at a tenth it is ten times above it.

2. A mud line 21 km downstream. Helicopter stills of the Upper Trishuli-1 headworks put the flood's mud line 45–70 m above the old bed, read four ways — a tunnel portal as a ruler, the debris benches below it, geopera's measured 11–18 m of bed rise, and the road ledge. The corridor map, on a better terrain model, then read 70–90 m in the same reach, so the stills were the low estimate. Runs that pass every envelope observable put 5–21 m of water there.

3. The mud-line map, which is the decisive one. On 7 September every mud line in the corridor was mapped from post-event imagery along both banks and the 8 m elevation model sampled beneath it (output/trimlines.csv, trimline_fit.csv). These are the water depths this analysis had never been scored against. Scored against them, the model wants a release nearer 110–175 Mm³.

Two independent lines published since make 110–175 Mm³ an ordinary number rather than an alarming one. The first academic reconstruction (arXiv 2609.04563) maps the source area at 1.009 km² and deliberately publishes no volume; over that area 110–175 Mm³ is a detachment slab 109–173 m thick, where Chamoli 2021 was ~135 m by the same arithmetic. And geopera's released volume of ~100 Mm³ ±40 % now overlaps the range from below. See §04b for the endpoint-angle cross-check from the same preprint.

13What would change my mind

This is v16, built thirteen days after the event on contested numbers, and five results have already been withdrawn or reversed (§00). It moves if the facts move.

A published source volume and ice fraction. The input that moves the most. At my own envelope the melt ceiling is 1.1–2.5 Mm³ (§04); it approaches the FFD's ~20 Mm³ only if the failed mass was predominantly ice and near 200 Mm³ and the heat partition sat at the top of the published range — the single combination that would put finding 01 in play.

The split of flow at the junction. The volume that ran up the Kyirong arm is measurable; the share that went that way is not, and the geometry cannot fix it (§12b). Anywhere from a tenth to a half, and the through-volume swings by a factor of five across that.

The DHM records at Betrawati (station 447) and Galchhi, both of which survived. A peak discharge at either settles the one question this model cannot: how much of the flood that passed Betrawati reached Galchhi, and how fast. I do not have them and have not asked for them.

A repeat elevation survey of the deposit. If it is ice-rich it must keep losing volume as buried ice melts out; rock does not. Two surveys a month apart would separate them with no model involved.

And the structural one, added 8 September. The lower-river inference is now being tested by sweeping the junction hydrograph directly — peak, duration, volume, composition — and asking whether any shape of flood at km 22 routes to the profile the mud lines describe (PLAN §11). If none does, the fault is in my reconstruction of those stages or in the lower-river physics, not in the avalanche, and finding 04 stays withdrawn rather than being resolved. That is a named outcome, not a hedge.

14Method & sources

River path stitched from OpenStreetMap waterways (Chhochen Khola → 东林藏布/Lhende → 吉隆藏布/Kyirong Tsangpo → भोटे कोशी → Trishuli), sampled every 400 m, elevations from the Mapzen terrain dataset via OpenTopoData. Energy and water budgets validated against Shugar et al. (2021). Front speed and peak celerity grid-fitted to the six timestamps. The complete workings — every equation, fit and figure, executed end-to-end from the data files — are published as a companion notebook: Trishuli Model Workings; a jargon-free telling of the whole analysis is The Water Was Already in the River. The interactive charts on this page re-compute the same ledger client-side. All code, data and the full evidence dossier (~30 sources, contested numbers flagged) live in the project repository. Analysis: Dave Hume, with Claude as research and modelling assistant.

• Timeline of the 2026 Nepal floods — Wikipedia
• 2026 Nepal Debris Avalanche and Flash Flood — USGS
• FFD technical report: 20 Mm³ excess surge — nepalnews
• Expert reaction to flash floods in Nepal — Science Media Centre
• The Landslide Blog, 26 August 2026 — Dave Petley, Eos
• DHM/FFD press release, 27 Aug 2026 (PDF) — official chronology & the 20 Mm³ method
• Valmikinagar barrage hourly discharge — Prabhat Khabar (Bihar WRD figures)
• How the Bhotekoshi flood unfolded over 10 hours — Kathmandu Post
• First peer-reviewed study (Chinese Academy of Sciences, in the journal Yangtze River, 人民长江) — via SCMP
• InSAR precursor creep before the collapse — Nature news
• FFD technical report, station by station — OnlineKhabar
• Bhote Koshi flood satellite reconstruction — Darcy Weedman, geopera (Geopera Pty Ltd company blog; not peer-reviewed)
• Rasuwa flash flood press release — ICIMOD
• Shugar et al. 2021, the Chamoli disaster — Science
• Seti River 2012 flood analysis (Kargel) — SANDRP
• Iverson et al. 2011, wet-bed entrainment — Nature Geoscience
• Huggel et al. 2005, the 2002 Kolka–Karmadon rock–ice avalanche — NHESS (the closest analogue, §04b)
• Haeberli et al. 2004, Kolka–Karmadon — Journal of Glaciology
• Kääb et al. 2021, glacier detachments and their reach angles — The Cryosphere
• Schneider, Huggel, Haeberli & Kaitna 2011, rock–ice avalanche mobility (64 events) — ESPL (abstract; full text not opened)
• Schneider, Kaitna et al. 2011, rotating-drum experiments on gravel–ice mixtures — Cold Regions Science & Technology (abstract)
• Sosio et al. 2012, runout of rock avalanches in volcanic and glacial terrain — Quaternary Science Reviews (abstract)
• Pudasaini & Krautblatter 2014, a two-phase rock–ice avalanche model — JGR Earth Surface (abstract)
• HiRISK rapid hazard assessment NP3, Rasuwa flood (28 Aug) — Steiner, Immerzeel, Allen et al. (PDF)
• China Ministry of Natural Resources reconstruction — Xinhua, 27 Aug (Chinese); the "nearly 3 km upstream" reading is Guo Zhaocheng's on CCTV, 28–29 Aug
• Road cut about 3 km from the port; Resuo village at 3 km reconnected — Xinhua, 28 Aug (Chinese)
• About 3 km of the G216 destroyed; the last 1.7 km buried — CCTV via China Daily, 4 Sept (Chinese)
• Planet disaster data for the event (PlanetScope, SkySat, Pelican; CC-BY-NC-4.0) — the sub-metre scenes read in §08
• Vantor (Maxar) open data collection for the event (CC-BY-NC-4.0)
• Tibet press conference transcript: barrier lake 2.2 Mm³, inflow ~23 m³/s — Xinhua, 30 Aug (Chinese)
• The border-gate clips blocked on Chinese platforms — China Digital Times, 30 Aug (Chinese)
• Cenderelli & Wohl 2001, GLOF attenuation — Geomorphology
• Pierson & Scott 1985, lahar dilution — Water Resources Research
• August 2026 Nepal–Tibet floods — Bethan Davies, AntarcticGlaciers (warming & retreat figures)
• The worrying climate link — Yale Climate Connections (the framing §11 responds to)