What changed, and when

9 October 2026 Paused for some months. Nothing new was computed. The findings stand as they were on 6 October: finding 01 holds on best evidence, finding 04 is the measured 110 to 125 million cubic metres, and composition is open. The first job on restarting is the test at the junction, and the things that would restart it sooner are listed on the front page. Corrections still go up here, paused or not.

6 October 2026 The collapse size goes into finding 04: 110 to 125 million cubic metres, measured, preliminary. It was computed on 5 October and has had its night, as the rule at the top of this page requires. The measurement is arithmetic on the elevation data David Shean and Shashank Bhushan released; if they publish a volume, theirs replaces mine.

The finding now says three things. The hole in the mountain is 110 to 125 million cubic metres, which is an upper bound for the 08:37 collapse because it counts everything that left the scar up to 6–8 September. The upper half of my model fits at that size, with a floor near 115 set by the border clock, so two routes that share nothing overlap at about 115 to 125. And 4 of the 15 runs that fit also land the downstream measurements I hold back. It also says what is still wrong: Galchhi fails in every run, 4 of 15 is thin, and nothing here says how much was ice.

What it replaces. Finding 04 was titled "How big the collapse was — unresolved, and may be withdrawn", with an amber chip. It said the September envelope of 14 to 34 was too small, that the mud-line map wanted 110 to 175 in the model, and that "this finding may end up withdrawn rather than corrected". Its chip is now blue. Every page was searched for 140 to 160, 110 to 175 and 14 to 34: where they were stated as current they are gone, and where they are the record they stay, dated.

5 October 2026 — afternoon Finding 01 is weaker than it was. I had the ice dial backwards, and at the measured size the corner reaches the flood's extra water. Carrying the measured 110 to 125 million cubic metres into the melt budget, I swept what the mass was made of instead of picking a value, and the answer went against me. Melt here is limited by the energy of the fall, not by the amount of ice, and rock is three times as dense as ice. So a rockier mass releases more heat, and the case most favourable to melt is a mass about a fifth ice, not four-fifths.

At the measured size, on best evidence, melt is 3 to 8 million cubic metres whatever the composition: less than half the Flood Forecasting Division's ~20 million. With the longest drop, the most generous heat-sharing published, and a mass a fifth to two-fifths ice, it reaches 21 to 24. I cannot rule that corner out. Whether a mostly-rock mass can put 70% of its heat into its ice, I have not examined.

What this replaces. Finding 01 said melt "caps out at 1.1 Mm³ on best evidence, or 2.5 with every dial turned in melt's favour", and reached 20 only "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". The first pair was computed at the 14 to 34 envelope ruled out on 1 October. The second has been wrong since 5 September: those inputs at a fifth ice give 38. This morning the technical report also said the measured size brought the absolute ceiling down to 11 to 13; that was the four-fifths ice row. The finding's chip goes from green to blue. It holds on best evidence, not across the range.

This goes up today because it weakens a finding; see the rule at the top of this page. The measured volume strengthens one, so it is still waiting for its night before it goes into finding 04.

5 October 2026 — later I ran the model again at the measured size. It fits, and better than at its own number. The trouble downstream is not the size. Dave's question was whether the smaller collapse would also fit further down the river. I wrote down the tests and what each answer would mean before running anything, and committed that to the repository first.

The top half. At the measured 110 to 125 million cubic metres, 15 runs of 300 match everything measured above Betrawati. At the model's old range of 135 to 200, on the same corrected geometry, 4 of 300 do. Across all 900 runs, the model fits best at 115 to 120, still fits up to about 150, and stops fitting below about 115, because a smaller mass reaches the border too late. So the measurement and the model now say the same thing from two directions: the hole in the mountain caps it at about 125, and the border clock puts a floor at about 115. The 140 to 160 I reported on 4 October was the bottom of the range I let the model try, not something it found.

Further down. Four of the 15 runs at the measured size also land the three measurements I hold back and never score — when the water reached Malekhu and Kalikhola, and the peak at Devghat — and the depth between Betrawati and Galchhi. That is the check that caught a false answer on 8 September, and the one I said splitting the model had switched off. It now passes, in 4 runs of 15.

Galchhi still fails in every run: 14 to 17 metres where the mud lines say 3.5 to 10. The smaller collapse barely moves it. So the problem at Galchhi was never that the collapse was too big. It is the lower river, where the flood spread across the valley floor in a way this model cannot, and that is now a cleaner question to ask.

Not in the findings yet. After a night it is a candidate to replace finding 04.

5 October 2026 I measured the collapse directly, and it is smaller than my model said: 110 to 125 million cubic metres. Preliminary, and not in the findings yet. I subtracted the mountain after from the mountain before, using the elevation data David Shean and Shashank Bhushan released on 18 September. The data is theirs and is licensed CC BY-NC 4.0. The arithmetic on it is straightforward, and they will very likely do it better. If they publish a volume, theirs replaces mine.

The scar shows up without any help: a slab cut out of a west-facing slope between about 4,300 and 5,200 metres, more than 200 metres deep in places, with the surface on either side unchanged to within a few metres. I did not borrow an outline. I let the change in height define it, and depending on where I draw the edge it covers 0.8 to 1.2 square kilometres, mostly 90 to 160 metres deep. The two published outlines are 1.0 and 2.0 square kilometres. Mine sits near the smaller one; the larger reaches about 300 metres further down the headwall.

Four tests were written down before I computed anything. At the threshold the plan named, all four pass. Ground that did not move reads zero to within 0.1 metre on average, with a scatter of under 2 metres. Two stricter thresholds I tried as checks fail two of the tests: the outline shrinks below half the larger published area, and over 40 per cent of it falls where the authors' quality filter has discarded data. Their volumes are inside the range anyway.

The glacier had thinned on its own between 2015 and 2026. I measured that on nearby glaciers the avalanche did not touch, at the same heights, and took it off: 2 to 7 million cubic metres, against the 6 to 12 I had guessed. I also repeated the whole measurement on two other pre-event surfaces made by different means, one of them radar, and two independent post-event surfaces. They all land in the same place.

What it does to my own number. The upper half of the model gave 140 to 160. The measurement is below that, and below the bottom of the range I let the model try, which started at 135. So the model was never run at the size the mountain says. That is not quite a failure of the model: it feeds its collapse in at the head of the valley, so its number should also include whatever the avalanche scoured off the headwall below the scar on the way down. I can see at least 20 million cubic metres taken off there and at least 7 put down, but half that ground has no data. So the measurement does not confirm 140 to 160, and it does not quite rule it out. It does say the model has to be run again with room to go lower.

And what it cannot tell me. It is everything that left the scar between the before pictures and 6–8 September, so it includes the smaller signals that morning and anything that fell afterwards. It says nothing about how much was ice. That still waits for next summer.

A correction that came out of it. The plain-English page said my volume, allowing for what the flow picked up, sat comfortably inside the seismic group's weight. It sat inside their outer bounds but below the spread of their own 24 attempts, and the measured volume sits lower still, just above their floor. The page now says so.

4 October 2026 — evening The pre- and post-event elevation data everyone said did not exist has been public since 18 September. And a way of weighing the mountain that nobody seems to have tried. The first academic reconstruction declined to publish a volume because "no suitable co-registered pre/post DEM pair exists". One does now: David Shean and Shashank Bhushan released pre-event composites (8 m from 2012–2016 and 2 m from 2015–2017) and post-event stereo at 2 m from imagery taken between 27 August and 8 September, plus radar stereo. No volume has been computed from it. The most direct measurement available — subtract the mountain afterwards from the mountain before — is sitting there undone.

Before downloading anything I asked whether the errors would swallow the answer, which is the only question that matters for a measurement like this. They would not, and the margin is large. Differencing two independent post-event surfaces over ground the avalanche never touched gives a bias of about half a metre to a metre and a scatter of around two. Against a scar of one to two square kilometres, a one-metre bias costs one to two million cubic metres on an answer near 150 million. It would take something like a hundred metres of systematic error to make the exercise pointless. The real difficulties are different: the pre-event surface is a decade old and the glacier has thinned on its own since, which is worth six to twelve million cubic metres and biases the answer high, and the steep shadowed terrain where the scar sits is where elevation data tends to be missing. The first of the released products turns out not to reach the detachment zone at all.

And a different idea, which is Dave's. The mountain was carrying that mass and now is not, so the rock beneath should spring back — a few centimetres to twenty, depending how fractured it is. The mass landed 40 kilometres downstream, so the ground under the debris should have been pushed down by a similar amount. Up at one end, down at the other, a known distance apart. It would also weigh the thing directly, because what the ground responds to is weight, with none of the guesses about density or acceleration that every method in use needs. Radar satellites have the precision; whether they can hold a usable signal on a steep snowy Himalayan face is another matter, and an earlier radar study of this slope found it could not. Written into what would change my mind rather than attempted.

4 October 2026 — later A claim of mine about how to settle the ice question, corrected. I have been saying for weeks that two satellite surveys of the debris pile a month apart would separate ice from rock, because buried ice melts and the pile sags while a rocky pile sits still. The test is right. The timing was not, and I had not done the arithmetic before repeating it.

The deposit sits at about 1,300 metres, low and warm, where bare ice would melt 3 to 6 cm a day — more than a metre since late August. But the ice is buried under rubble, and rubble insulates it well: half a metre of cover cuts the melting by five to ten times. So the sag to date is more like 0.1 to 0.5 m, against a satellite stereo precision of about 0.5 to 1 m. Marginal to invisible. October is also the wrong end of the melt season.

The useful comparison is across a full warm season, roughly May to September 2027, by which point an ice-rich pile should have dropped metres and a rocky one centimetres. So the composition question stays open for most of a year, and the plain-English page now explains the whole test properly rather than asserting it.

4 October 2026 The top half of the model fits. Saying so, with the reason it is not a finding. Splitting the problem at the border on 8 September left an upper half that can be tested on its own, and it passes: six runs of 300 match everything measured above Betrawati at once, putting the collapse at 140 to 160 million cubic metres released over four and a half to seven and a half minutes. That is the first time any version of this model has satisfied a complete set of tests.

It is not going into the findings, and the reason is structural. The way I caught the last false answer was by holding three downstream measurements back and checking against them afterwards. All three sit below the border. Splitting the model at the border switched off the only machinery that has ever stopped me believing a bad fit. Six of 300 is thin, one input is not constrained at all, and finding 04 stays under revision.

Composition is still not resolved, and now I can say exactly why. The model leans towards a lot of ice, and that leaning rests on a single measurement: about 12 million cubic metres of rock left in the valley. Relax that one constraint and the middle of the range falls from about 90 per cent ice to about 45, with over half the surviving versions carrying less ice than rock. The 12 million is a lower bound from less than half the valley floor. Marc Hendrickx, working the same reach independently, uses 13 per cent ice and gets a working model; so do I, if I relax that number. Two surveys of the deposit a month apart would settle it, because buried ice keeps melting out and rock does not.

The plain-English page now carries all of this, along with an account of the September research written for a general reader — how a group weighed the avalanche by measuring the shove it gave the Earth, why that weight does not contradict mine, and how the same work caught me claiming the front never slowed down.

1 October 2026 Three weeks of other people's work, one correction to mine, and a seismic estimate of the mass that does not disagree with ours after all. I had not looked since 12 September. Quite a lot has appeared.

The correction first. These pages said the flood front did not slow down over the first 22 km. A near-field seismic analysis — 16 regional stations, pattern correlation, Huang, Wang and Chen at the University of Southern California, manuscript in preparation for Geophysical Research Letters — breaks the run into legs and finds the front at 76.5 m/s over the first 1.4 km, 111 m/s to the big bend at kilometre 6.6, then 38.8 m/s over the last 14.3 km. It accelerated off the headwall and then slowed through the gorge. The mean over the whole run is unchanged and agrees with mine closely — their front reaches the border at 458 seconds against my 460.6 — but the shape inside it is not what I said, and the claim is now off the pages. It does not sit comfortably with my own instantaneous reading of about 47 m/s at the border either, and I have recorded that disagreement rather than resolve it.

The collapse happened in two seismic stages, 22 seconds apart. The same work resolves an initial detachment and then, 22 seconds later, what it calls the massive runout — the mass reaching the end of the headwall 1.4 km down and hitting the valley floor. That impact is what radiated the magnitude 5.2. Their checkpoint at the border post sits 56 metres from mine, which is the closest independent check on my geometry so far.

And a mass, from seismology. The same group inverts the force the avalanche exerted on the Earth from long-period surface waves at 47 stations worldwide and converts peak force to mass. Their range is 0.53 to 2.04 × 10¹² kg across 24 variants of their own analysis, with stated bounds a factor of nine wide. I first read that as disagreeing with my 140–160 million cubic metres. It does not. Force inversion measures the mass that was moving, 7 to 21 kilometres down the valley, after it had entrained everything it scoured; my number is what detached at the crown. Bulked by the 33 per cent another analysis uses, mine lands inside their range and above their floor. What their result does do is close the bottom of the argument: this cannot have been the 14–34 million cubic metres I published in early September, and the seismology rules that out by an order of magnitude.

Also since 12 September: a three-phase rock–ice–water simulation of the event sweeping releases of 200, 300 and 400 million cubic metres with melting switched on and off; a climate attribution study putting record melt-season warmth as a precondition; a multi-institution field and imagery assessment led from Hirosaki University; and a second observational reconstruction whose arrival time at the first gauge, 12.8 minutes, is within a fifth of a minute of the 13.0 my model was never fitted to. Links are in the technical report. Nothing here changes a finding except the one retracted above.

8 September 2026 — evening The seventh version failed too, and we have stopped building versions. A measurement of our own scoring is the reason. The floodplain version ran: four hundred combinations, none passing. It did what the video said it should — the depth at Galchhi is now right in seventy runs where it was fifty before, and the flood covers the valley floor at about the rate the clip shows — but the failure split in two rather than closing. The dry runs now reach Galchhi at the right depth and the town below it on time, which no previous version managed, and pay for it by leaving four times more rock behind than satellites measured. The wet runs leave the right amount of rock and stand five metres too deep at Galchhi.

What stopped the work was not that. It was a check we had never run: comparing the model's water depth against the mud lines at every one of the eleven hundred points along the river, rather than as six reach averages. The one run that has ever passed all eleven tests matches the mud lines at eleven per cent of the points in the first twenty-two kilometres. It satisfies every average while being wrong at most places inside them. Six of the eleven tests are reach averages of a profile we reconstructed ourselves, so the scoring was too coarse to pin the model down — and seven versions of failure look less like a search that has not converged and more like a question that was not answerable as posed.

So the model is being split in two, which is Dave's idea and the right one. The first twenty-two kilometres are a rock-and-ice avalanche in a confined gorge, with almost no river to sweep up and a hard measurement to hit — the border camera's seven minutes forty. Everything below the junction is a flood wave on a monsoon river, which does not carry its own water but gathers the river's as it goes. Those are different problems with different physics, and joining them at a single handover — the shape of the flood as it passes the junction — lets each be tested on its own instead of one being blamed for the other's failure. The first experiment needs no avalanche model at all: is there any shape of flood at the junction that can be routed down the river we measured? If the answer is no, that is a real result about our reconstruction or about the lower river's physics, and we will publish it as one. Finding 04 stays under revision, and may end up withdrawn rather than resolved. Plan §11; dossier §27–28.

8 September 2026 — midday The first academic reconstruction of this collapse has appeared, it agrees with our geometry, and its source area makes our volume an ordinary one. And a figure we went chasing turned into a correction of ours. A preprint posted this month (arXiv 2609.04563) reconstructs the collapse from open data. It deliberately publishes no volume, no velocities and no dynamic model, on the stated grounds that one would add false precision without a source volume, material fractions and a hydrograph. Two of its numbers bear on us. It measures the straight line from where the mountain broke away to where the debris stopped at 8.84° below horizontal; we had 8.76°, from our own map, path and choice of crown and toe. That ratio is a measurement of the physics — for anything sliding to rest it is the friction — so the agreement says our geometry is sound, which matters because every speed here is a distance over a time along that same line. It is a simple measurement on similar public data, so agreement is close to what one should expect: it validates the foundation, not the building, and nothing about our volume or our model. Second, it maps the source area at 1.009 square kilometres. Spread our 110–175 million cubic metres over that and the slab that detached was 109–173 metres thick; Chamoli 2021 was about 135 by the same arithmetic. So the number that has read as uncomfortably large all week is, as a piece of mountain, unremarkable — the first independent constraint pointing that way, and it arrives from a study that took care not to estimate a volume at all. Detail, including the boulder-size velocity check that agrees with our gorge window, is in report §04b.

The correction. Our literature search picked up a figure — an "additional flood volume" of 19.96 million cubic metres — and for a few hours we treated it as a new constraint on the size of the collapse. Then we went to the source, and found it already in our own data folder, downloaded on 2 September: a two-page press release in Nepali from the Flood Forecasting Division, issued the day after the flood. It says the volume above base flow measured at Devghat, between 14:10 and 18:00, is preliminarily estimated at approximately 2 crore cubic metres — 20 million, one significant figure. That is the extra water past one gauge inside a stated four-hour window, not a measurement of the whole flood, and it cannot be set against a release volume. The two-decimal version we had been working from is not in it. Our technical report has had this right since it was first published; the mistake was ours, made by working from a summary of a document we already held. We have transcribed and translated the release in full so anyone can read it rather than take our word for it. What the figure can do is fix how wide the wave was at Devghat, which no version of this model has ever been tested on, and that is now a test.

The same release records a magnitude 2.6 event near the border at 08:20 and the main one at 08:37 — a precursor seventeen minutes ahead. Chinese researchers have separately reported unusual signals at about 06:04, 06:50 and 08:15. Two independent records now point at a mountain that may have failed in stages, where every version of our model has assumed a single release. Nothing here changes a finding. Dossier §26.

8 September 2026 — early morning The first run to pass every test — and why it is not the answer. Then a video that shows what the answer is. A second-stage sample inside the region the night's best runs bracketed produced, for the first time, one run of three hundred that meets all eleven tests. Our rule is that every passing run is then checked against three measurements it was never scored on, the arrival times at Malekhu and Kalikhola and the peak at Devghat, and this one fails all three: forty minutes late, two and a half hours late, and half the peak. It is a hundred and sixty million cubic metres with almost no water in it, a stiff, slow mass that reaches Galchhi at the right depth by crawling. The runs that get the downstream timing right are the wetter ones, and every one of them is too deep at Galchhi. So the model can be right at Galchhi or right below it, never both, and that is a statement about what the model lacks, not about the flood. Dave found the missing piece on YouTube: a clip shot from a hillside above a broad farmed valley on the Trishuli. At the start the river is a braided gravel channel perhaps a hundred and fifty metres wide. Seventy seconds after the front appears, the flow has left the channel and covers the entire valley floor, several hundred metres across, to the houses at its edge, and it stays there, moving slowly, for the rest of the clip. Our mud-line map has the same event in it, unread until tonight: the water depth halves at kilometre ninety, exactly where the elevation model shows the valley floor opening from two hundred and fifty metres to six hundred. A flood that spreads over a floor like that becomes long, low and slow at once, which is what the record downstream shows and what a single channel cannot do. The next version gives the lower river its floor, with the water allowed to move across it, and is argued for in dossier §25. The location of the video is not yet pinned; a large mid-channel boulder and a cable crossing in the first frames should fix it. Six versions, one pass, none promoted. Finding 04 unchanged: under revision, moving up. Dossier §24–25; output/ensemble_v10b_RESULTS.md.

7 September 2026 — night Fifth version, three hundred runs, none pass — and the upper corridor is now solved in the model. What is left is the lower river. Dave asked whether our limits were wrong, or whether there was a non-linearity we did not understand. Both, as it turned out, and the answer is in dossier §23. Three limits were wrong and none of them were being sampled: the collapse had been fed into the valley over a fixed three minutes in every version, an assumption standing in for a seismic duration nobody has published, and it acts directly on the border clock; the dry-friction floor sat above the values ice gives; and the deposit limit is a measurement of rock, from elevation models made days after the event, that had been applied to ice as well. With those three moved — release duration sampled up to ten minutes, ice-capable friction, deposition scored on rock alone — fourteen runs now meet the border clock and the gorge depth together where the previous version managed one, three runs meet ten of the eleven tests, and the two best fail on exactly one thing: too much water at Galchhi, eighty-five kilometres downstream. The non-linearity is that every limit on the flood's speed in this model grows with the square root of its depth, so volume, valley width, composition, drag, deposition and the clock are one coupled system rather than six dials, and the search had been walking its boundary on assumptions it was never allowed to vary. The remaining conflict discriminates cleanly: wet releases put fourteen to nineteen metres at Galchhi and arrive at Malekhu half an hour to an hour early, dry ones put one to seven metres there and arrive late or not at all. The answer is between them, and a single 267-metre channel cannot produce it in a valley the site table puts at five hundred. Next: the lower river's floodplain, from the elevation model, as the gorge and the Lhende were done before it. Where the model is converging, stated as the shape of a search and not a finding: 110 to 175 million cubic metres, mostly ice and water, released over five to ten minutes. Finding 04 unchanged: under revision, moving up. Dossier §23; output/ensemble_v10_RESULTS.md.

7 September 2026 — evening A fourth version, and the failure has moved again — this time onto a measurement, not the model. The afternoon's result was that one drag setting could not serve the gorge and the lower river at once. So the drag was made to follow the flow's composition, which the model already tracks: full drag for a rock-and-ice mass, none for water, and a mud-rich flow keeps its drag because it is still a debris flow. Zero of two hundred again, and the reason is instructive. The runs that carry the drag are the ones that start with little water in them, and in this model such a flow lays down fifty to sixty million cubic metres of solids along the corridor, against a measured twelve, and then stalls in the flatter river below Betrawati. The wet runs pass the deposit limit and keep Galchhi in range, but have no drag, and are back to running the gorge like water. One run is worth naming: a hundred and thirty million cubic metres, three-quarters water at the moment of detachment, with a wide upper valley and weak drag. It matches every measured depth except Galchhi, the gorge speed, the scour volume, and the three downstream numbers it was never scored on — the fronts at Malekhu and Kalikhola and the peak at Devghat — within five to fifteen per cent. It fails the border clock by six seconds, Galchhi by six metres, and the deposit limit by ten million cubic metres. Not a pass; the shape of one. Which points at the limit itself. The model counts unmelted ice as solid, and the twelve million cubic metres comes from elevation models made three to six days after the event, by which time any ice laid down in the corridor had melted. The limit is on rock, and it has been applied to the ice as well. Next, for a decision: score deposition on rock alone. Four versions, none passing, all pointing the same way. Finding 04 unchanged: under revision, moving up. Dossier §22; output/ensemble_v9_RESULTS.md.

7 September 2026 — afternoon Two more versions of the size test, and the failure has moved to a more interesting place. The midday result said a release deep enough to leave the mud lines runs the gorge as a frictionless bore. So the model was given a turbulent-drag term of the kind used for rock–ice avalanches, as a sampled input alongside the others, and the loss at the border junction was fitted rather than assumed. On the deepest near-miss that fixed the gorge speed at once and did nothing to the clock — and the reason turned out to be above the junction, not in it. The model's Lhende valley was 50 metres wide, a rule of thumb; the mud-line map has it at 95 to 190 metres over the ten kilometres above the border. In a 50 m channel a hundred-million-cubic-metre release runs two hundred metres thick, and at that depth every limit on its speed scales with the square root of the depth. So a third version added the mapped width as a sampled input. Both versions still pass nothing — zero of two hundred, twice — but the pair the midday run could never produce at any size, the mapped gorge depth and the border clock together, now exists: six runs, all a hundred million cubic metres or more. What one drag coefficient cannot do is serve the gorge and the lower river at once. Sized for a seventy-metre debris flow it stalls the muddy flood below Betrawati; sized for the lower river it leaves the gorge too fast. The nearest miss, a hundred million cubic metres with the gorge's drag, reaches Malekhu and Kalikhola within five to ten per cent of the observed times without being scored on them, and fails only by putting sixteen metres of water at Galchhi where the gauge saw nine. Next, and it is Dave's decision, not the session's: make the drag depend on how much rock the flow is carrying, which the model already tracks, so the gorge and the lower river stop asking for different physics. Finding 04 is unchanged: under revision, moving up. Also found and logged: a text-versus-number slip in the width filter of the midday run that kept the sharpest imagery out of it; it moves widths by less than a third and nothing in the midday diagnosis depended on it. Dossier §21; output/ensemble_v7_RESULTS.md and _v8_.

7 September 2026 — midday The size envelope was re-run against the mud-line map, and nothing passes — which is the most useful result this project has had. Two hundred releases from one to two hundred million cubic metres, scored on the corridor's mapped water depths for the first time, with the measured channel widths, the Chinese arm as the pond the map found, the corrected clock and a speed compared like with like. Not one satisfies the depths and the border clock together, at any size. The runs deep enough to leave the mud lines are 86–142 million cubic metres, three to five times the envelope on this page — and every one of them reaches the border in 4.2–4.8 minutes at 45–62 m/s where the record says 7 min 40 s and about 34. The runs that make the clock never put more than 31 m of water in a gorge whose mud lines stand at 60 to 100. Roughness does not bridge it: at seventy metres of depth the model's friction all but vanishes and a big release runs as a water bore at the Froude limit, while the real flow was that deep and half as fast. So the structure is wrong, not the numbers: a debris flow this deep carries a resistance the composition dial does not have. The measured widths, for the record, are one and a half to three times wider than the rule of thumb, which kills the long-standing excuse for the low stages. Next: add a turbulent-drag coefficient to the friction law as a seventh sampled input and re-score, because the volume and the rheology now have to come out together. Finding 04 stays where it is: under revision, moving up. Dossier §20; output/ensemble_v6_RESULTS.md.

7 September 2026 — morning Every mud line in the corridor, mapped from imagery plus the terrain. The stripped-ground boundary was extracted from Sentinel-2 and from Planet's 0.55 m Pelican scenes along both banks at every 100 m from the scar to Betrawati and 6 km up the Chinese arm, and sampled on the 8 m High Mountain Asia DEM (fetched this morning; the 30 m DEM put geopera's 134 m cliff 90 m lower at the same point). It was gated first at the border: the method returns the three mud lines Dave measured there — bed 1,815, lee line 1,875, impact cliff 1,920–1,930 m — as 1,814–1,818, 1,866–1,878 and 1,917 m, each ±5 m, and geopera's 134 m maximum turns out to be the same wall. Then Hakubesi, where it reads ~85 m above the pre-event bed where the helicopter stills had given 45–70 and geopera's own points give 75–107, so the stills were the low read and the passing runs' 5–21 m are 4–17× short. Then Syabrubesi, where the velocity collapse geopera reported reproduces as 12–15 m/s at their station. Along the whole gorge the stage is 62–72 m median with both banks valid at 127 of 128 stations; bend pairs give 37–47 m/s where they pass a Froude and error grade, and peak discharges of order 300–400 thousand m³/s. The one thing that moved: up the Chinese arm the mud line is level at ~1,875 m for 2.5 km — the lee line, not the 1,925 m contour the up-valley volume route used — and a pond at that level holds 23 Mm³ (15–33), with the widths the sub-metre imagery showed; the 170–690 Mm³ corners of that route go with it, and its best corner now sits inside or at the top of the 13–34 envelope. Coverage, flags, per-point errors and the CC-BY-NC handling are in dossier §19 and output/trimlines.csv. Nothing promoted: the profile is the observable set for the ensemble rerun, which is the next job.

Late morning: the whole 200 km, and a fit. Dave looked at the junction plan view with the Pelican image under the contours and called the highest points on the Nepal-side wall as cloud shadow; they are now a named override, and the image-plus-contours plot is a tool anyone can run on any box. The map was then extended to Devghat with a second Sentinel-2 mosaic and a third 8 m tile, and a robust running fit (±1 km median, outliers beyond 2.5 MAD marked rather than averaged) put through 1,228 bank points, rejecting 80. Stage falls from 73 m in the gorge to 29 above Betrawati, 15 between Betrawati and Galchhi, and 7.4 m [4.7–9.9] at Galchhi against the gauge's +9 m. Below Mugling the method mostly finds the pre-event gravel margin rather than a mud line — 606 of 665 bank points show no change — so the last 50 km carry only 56 clean points and are reported as such. Dossier §19a; output/trimline_fit.csv.

6 September 2026 — night A third cold read, this one aimed at the dependencies rather than the prose, and it found that this morning's correction never reached the code. The 6 September clock fix — 6 min 50 s to 7 min 40 s — was applied to the hero stats, the clocks table, the findings and the plain-English page, and to no model constant at all. Six code files were still scoring against 7.0 minutes, which means the 14–34 Mm³ envelope had been computed against a number this site had publicly retracted eight hours earlier. Fixed in all six and both ensembles rerun. The envelope barely moves — 13.3–34.0 Mm³, median 23.2, from 21 of 200 — and the smallness of that change is the reassuring part, not the interesting one. The interesting part is that a retraction is not finished when the pages read right.

The CAS re-score, promised on three pages for five days, was run. Scored against the peer-reviewed 19 m/s instead of the 45–52 superelevation band, nothing survives — 0 of 200. The conflict is not about size: no release anywhere between 1 and 200 Mm³ can satisfy the border arrival and a 19 m/s border speed at once, and none can satisfy the Syabrubesi arrival with it either. Each is reachable on its own. That is consistent with our argument that CAS measures the post-turn water surface rather than the front, and equally consistent with our speed-to-arrival physics being wrong; it does not choose between them. Both scorings are now published rather than one asserted.

The profile-integrity check the Seti post-mortem called for on 3 September had never been built. It exists now, it gates rather than warns, and it audits the profile the model actually consumes rather than the intermediate — a distinction that matters, because its first draft reported the live Trishuli channel as failing and was itself wrong. Seti as published fails all four gates; the repaired Seti channel still fails one. The Trishuli lower river does turn into a staircase under the monotone clamp — 82% of the last 39 km at zero gradient — but the smoothing that follows removes it, so the elevations are cleared and the Galchhi stage failure now points squarely at the channel widths, which below km 60 are a rule of thumb rather than a measurement.

On the Chinese arm, from the routing side. Modelling the Kyirong arm as a backwater wedge rather than a linear store and sweeping release volume against wetness, no combination reproduces the Galchhi rise and the border clock together. The nearest miss — 60 Mm³ at 40% liquid — clears Galchhi at 10.6 m and arrives at the border eight seconds too early. Under this morning's superseded 7.0-minute clock it would have passed. That is a structural result about the store, not a size, and it points the same way as the up-valley volume route above: up. Nothing promoted.

Also on this page: the confidence chips now have a legend, the FFD's 20 Mm³ is cited where it is used rather than referred to, the not-peer-reviewed disclosure has moved from the last finding to above the first one, geopera's independence is described as what it is — a different solver over shared inputs — and finding 02 now says which end of its seven-minute clock has actually been audited. It is the camera. The seismic origin at the other end is a catalogue solution taken on trust.

6 September 2026 — late A new, loose line of evidence on the collapse size, and it leans larger. Dave Hume's method: treat the border junction as a node, measure the volume that ran up the Chinese arm from the valley floor and the imagery, split it at the node by the valley geometry, and subtract what the Lhende contributed. On the Copernicus 30 m DEM the arm held somewhere between 13 and 70 million cubic metres at peak depending on how the mud-line head is read. The share that went up the arm is the crux, and the valley geometry cannot fix it: the Lhende's last 500 metres bend toward the downstream exit, the kilometre above runs at right angles to both exits, and a 100-metre-deep flow at 170 km/h does not follow a 500-metre bend. Anywhere from a tenth to a half is defensible. Divide one by the other and the corners straddle finding 04's envelope: one lands inside it, the next at its top, and most well above. It is a first pass with an unmeasured split and a 30-metre DEM, so it cannot move the finding, but it leans larger and it is recorded here rather than held back. What would fix the split is peak-discharge continuity at the junction — mud lines and superelevation just above and just below it, which helicopter footage can supply — and the tool for that is built. Working in the dossier (§17); what would tighten it is listed in report §13.

Later the same night, a second line pointing the same way. Helicopter stills of the Upper Trishuli-1 headworks reach, 21 km below the junction, show a mud line roughly 45–70 m above the pre-event river bed — read four ways without imagery: a tunnel portal as a ruler, the deposit benches, geopera's measured bed rise of 11–18 m there, and the road ledge from the map and the DEM. The runs that pass our size envelope put 5 to 21 m of flow at that spot. No stage anywhere in the corridor has ever been one of the observables the envelope was scored against; the next version will score them, and the envelope is likely to move up when it does. Dossier §18.

6 September 2026 — evening A second cold read, an account of the event, and a rethink of the first 22 kilometres. The technical report now opens with a description of what happened — what fell, from where, onto what, what the border camera saw, where the deposits are, what the mud lines record, and what nobody yet knows — and the plain-English page carries the same account in plain words. The video measurement listed below is written up in full on both pages: the clock, the two-route camera calibration, the two sight lines 50 m apart that the front crossed in 1.066 s, the error budget and what failed, with a plan-view diagram of the geometry.

The up-valley question has an answer. Re-measuring the Sentinel-2 change along the river centreline instead of in wedges, the ground stripped up the Chinese arm fades out between about 2 and 3.5 km from the junction, while the two arms the flood flowed through stay stripped for the whole 5–8 km in view. With Xinhua's "about 3 km" of damaged highway the 3.5 km contour trace, and — found in tonight's literature check — China's Ministry of Natural Resources' own statement that part of the flow ran "nearly 3 km upstream in the Gyirong Tsangpo", four lines now put the surge about 3 km up the valley toward Gyirong. Sub-metre imagery shows fresh sediment to about 4.6 km, which may be ponding behind the junction plug rather than the surge; the report keeps both. Report §08.

Dave Hume asked whether the first 22 km carried much water at all. They did not: the Lhende held about half a million cubic metres against a collapse of 14–34 million, so what struck the border was at least four-fifths mountain — a rock–ice avalanche, not yet a flood. The run down the Lhende is what an ice-rich mass does on its own, and the physics is the physics of an ice skate: a premelted film at rest, and at avalanche speeds frictional melting that renews a lubricating film at millimetres per second for a few per cent of the fall's energy — lubrication, not water production. Kolka–Karmadon 2002 covered a similar distance in a similar time. This is written up as report §04b and on the plain page; it reframes finding 02 without moving it, and it locates finding 01 below the junction, where the 22 Mm³ of river was. Not a scored result, so it is here and not in the findings.

Corrections from the cold read: §01 of the report still said "six minutes and fifty seconds"; the plain page's cliff box still compared the mud-line speed with the old 52 m/s clock and called it a match "within a few per cent" (it is within ~15% of the corrected 47.8); its opening still carried a toll ten days old; "only something running like slush" and "wall of water" both claimed more than the evidence shows; the version numbers disagreed between pages. And the word "survivors", which this site had been using for the model runs that pass its tests, is gone. It should never have been used that way here.

6 September 2026 We stopped taking anyone's word for the border camera, and it cost us a number. Until today the timestamp that the composition finding rests on reached us only through a partisan outlet's account of what the overlay said. We obtained the footage and read it directly: the station's own clock is legible and tracks real time. That retires the awkward part of the argument — and corrects our own figure. The destruction begins at 10:59:50 Beijing, 08:44:50 in Nepal, which is 7 min 40 s after the collapse and not the 6 min 50 s we had published. Ours came from reading the overlay to the minute and assuming zero seconds. The 22-kilometre average front speed drops from 53.7 to 47.8 m/s — 12% slower. The finding stands: seventeen and twenty-three minutes are still excluded, the corrected 7 min 40 s lies inside the ±30% tolerance the composition test used, and the Syabrubesi gauge still carries the argument without any camera at all. The runs themselves had not been repeated at the corrected clock when this was written; they were, later that night — see the entry above.

Also corrected: this site claimed the contact stress between ice grains collapses the melting point by about 70 °C. It does not. That was a straight-line extrapolation past the point where ordinary ice stops existing; the real floor is about 22 °C, and beyond it the high-pressure forms of ice melt hotter, not colder. The mechanism is premelting — a liquid-like film that is there without being squeezed — which is why you can skate at −30 °C. The conclusion it supported is unchanged.

Not promoted to the findings above, per the rule below: a first direct measurement of the flood front's speed at the border, taken off the video against two surveyed sight lines. 47 m/s, envelope 41–63. It agrees with the corrected clock and with geopera's superelevation, and it excludes the one published figure for the post-turn water — which is evidence that those two numbers were never measuring the same thing. It is a day old and needs checking; the method is now written up in report §03 and on the plain page.

5 September 2026 — evening A cold read of the whole site, and the corrections it produced. Three of them matter. The melt ceiling was quoted as sitting "far below" the official 20 Mm³ while this site's own chart showed 15 against 20; carrying finding 04's size envelope into the energy budget — which should have happened the moment that envelope existed — puts the real ceiling at 1.1–2.5 Mm³ and makes the claim both true and much stronger. A bar on the plain-English chart labelled "river already lying in the channel, 47 Mm³" was the total wave at Devghat; the standing channel water is 22.4. And the border-clock resolution was presented with more confidence than its sources carry: the decisive overlay quote comes from a partisan outlet and Kargel's matching figure has no stated method, both of which are now on the page. Also: the published workings had fallen three days behind the findings and have been rebuilt; every page now says who geopera is; and the human toll is at the top of this page, where it should have been from the first version.

A rule, adopted here rather than just intended: nothing reaches this findings list on the day it is computed. New results go in this changelog first and move up only after they have survived a night and one out-of-sample check. Findings 04 and 05 were promoted the same day they were produced, which is how finding 03 got promoted and then withdrawn.

5 September 2026 Two new findings. The collapse was probably 14–34 million cubic metres, an order of magnitude below the largest published estimates. And the model was proved structurally wrong before it was fixed — no combination of inputs could satisfy the timing and the debris budget together until sediment was split into coarse and fine. A challenge to the border clock was chased down and resolved in the camera's favour. The Nepali translation was withdrawn pending review by a Nepali speaker.

3 September 2026 — night The Seti 2012 result is withdrawn. Building the missing erosion physics meant looking hard at channel gradients, and that exposed a bad channel profile underneath the Seti test: 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. Nothing about the Trishuli findings depends on it. Meanwhile the erosion term now exists, and it puts the amount of material the flood tore out of the Langtang corridor within range of what satellite stereo actually measured — while saying the corridor cannot have received anything like the amount of rock the largest source estimates imply.

3 September 2026 — evening A deposition figure that geopera had themselves retracted was corrected in our scoring, which resolved a tension between the timing and deposit evidence — both now point the same way.

3 September 2026 The Chamoli 2021 timing gap closed by adding thermal lag to the melt physics. The corridor map was redrawn with automatic label placement.

2 September 2026 First public version: the water budget, the snowplow model and the river-as-circuit routing model.