Trishuli 2026 — Model Workings¶

Project home · Plain English · Technical report · Model workings · source and revision history on GitHub

What this is. The executable workings behind our analysis of the 26 August 2026 Langtang Lirung ice-rock avalanche → Lhende Khola / Bhote Koshi / Trishuli flood: where the water came from, and how the flood pulse evolved over ~200 km of channel. Five model stages run top to bottom, in the order they were built:

1. Energy & water budget — independent bounds on each candidate water source (frictional ice melt vs river-derived water), validated against Chamoli 2021.
2. Snowplow-integral model — 1D front-routing along the DEM river profile with per-source water bookkeeping, calibrated to the observed arrival clocks.
3. Ladder network — a 90-node routing model as a nonlinear R(L)C transmission line (diffusive-wave v0, upgraded to the local-inertia RLC form) for pulse structure: attenuation, side-valley charging, and the signature a temporary dam breach would leave.
4. Entrainment, and the falsification of the single-phase model — the erosion term the first three stages lacked, the rejection sampling that then proved the model form wrong, and the two-phase fix.
5. The consistency envelope — which combinations of release volume and composition are simultaneously consistent with everything measurable, giving 14–34 Mm³.

Dates. Sections 1–3 are dated 2 September 2026 (event + 7 days), with all inputs stamped to what was public then. Sections 4 and 5 were built on 3–5 September and added to this notebook on 5 September. Until then this page was published three days behind the findings it was supposed to make checkable — the summary pages announced results whose workings were not here. That was the wrong way round, and the fix is this notebook rather than a footnote about it.

Data files (in data/):

• river_profile.csv — 499-point channel profile, collapse scar → Devghat (199.3 km), from an OSM-stitched river path with Mapzen terrain-tile elevations (built by model/build_profile.py + model/fetch_elevations.py)
• valmikinagar_barrage.csv — Gandak barrage hourly releases at the India border (press transcriptions), downstream routing context in the full report
• ffd_report.pdf — official DHM/FFD press release (27 Aug): station clocks, Devghat peak 5,850 m³/s at 16:00, and the "~20 Mm³ excess water" figure

Authorship: Dave Hume, with Claude as research and modelling assistant.

Honesty rail. The two most important inputs are not published: the source volume (public estimates span 0.5–200 Mm³) and the ice fraction of the collapsed mass (unknown). Every conclusion below is therefore an envelope over defensible inputs, not a point claim. FFD's "~20 Mm³ excess" is single-source with an unpublished method; geopera velocity/height numbers are provisional. Where the models disagree with observation, we say so.

The border clock in this notebook. The border arrival is scored at 7.68 min (7 min 40 s, read directly off the CCTV overlay on 6 September). Until late on 6 September this notebook — and every model constant — still carried 7.0 min while the pages already said 7:40; that was caught in a third cold read and fixed in six files, and the ensemble was rerun (13.3–34.0 Mm³, median 23.2, 21 of 200, against 14–34 / median 21 from the 7.0-min run). The speed panel in Section 2 still plots the 52 m/s point under its original label ("Kargel, 193 km/h"); the corrected 22-km mean is 47.8 m/s, and the ensemble's speed observable is geopera's 48.5 ± 35%, unchanged.

1. Energy & water budget¶

The physics in one line: melt is energy-limited. Melting ice takes latent heat L_f = 334 kJ/kg. The only energy source is the fall itself: potential energy mgh gives 9.81 J per kg per metre of drop. A parcel falling the ~1,200 m from the Langtang Lirung scar to the channel carries ~12 kJ/kg — about 3.5% of what it would take to melt its own mass of ice. Even over a full ~2,400–3,500 m drop, melting a Chamoli-like ~8% ice mass fraction needs most of the dissipated heat to enter the ice, and realistic heat-to-ice partitions in the literature are tuned (0.3–0.7), not measured.

So we bound each candidate water source independently:

• A. frictional melt = min(energy cap, ice-mass cap)
• B. standing river water in the channel ahead of the wave (swept up)
• C. river inflow during the event's passage
• D. pore water in entrained saturated channel sediment

Code adapted from calcs/energy_water_budget.py — identical numbers.

Validation: Chamoli 2021¶

Before pointing this at Nepal, the check. Shugar et al. (2021, Science) reconstructed the Chamoli disaster as 26.9 Mm³ at ~80:20 rock:ice by volume, falling ~3,400 m over a short 26 km winter runout, with near-complete melt of the ~5–6 Mm³ of ice; their energy balance implies ~80% of dissipated heat entered the ice ("almost exactly the critical value required for near-complete melting"). River terms were near zero (winter low flow). The scenario function should reproduce ~5 Mm³ of melt, hitting the ice-mass limit rather than the energy limit — Chamoli is the rare case with energy to spare.

It does: melt water 4.8 Mm³ (~5 Mm³), and it is ice-limited — the energy cap (5.0 Mm³) sits just above the ice cap, Shugar's "critical value" in our units. River-derived share is ~8%. That freak 80:20 melt-dominated outcome is what made "melted glacier ice" the default narrative for 2026.

Langtang 26 Aug 2026 scenarios¶

Now the same arithmetic under monsoon flow + a much smaller initial fall + a 100–199 km runout, over a range of contested source volumes and ice fractions.

Scenarios 1–5 are as written on 2 September, when the published source volumes were the only guide and scenario 5 (~100 Mm³, per Kargel 50–200 and Azam/ICIMOD 100–200) was the best evidence available. Scenarios 6–8 were added on 5 September, carrying Section 5's consistency envelope — 14–34 Mm³, median 21 — back into the energy budget. That should have happened the day the envelope existed: until it did, the summary pages were quoting a melt ceiling of 15 Mm³ computed for a collapse five times larger than our own evidence supports, and describing it as "far below" the FFD's 20 Mm³, which it is not. Both rows are kept so the correction is visible rather than tidied away.

Scenario 8 is the one that matters for the argument. It is the melt narrative steel-manned until it breaks: the largest published volume, 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.

Budget verdict. Read scenarios 6 and 7 first: at the collapse size our own evidence supports, frictional melt is 1.1 Mm³, or 2.5 Mm³ with every assumption turned in melting's favour — short of the FFD's ~20 Mm³ of "excess" water by a factor of eight. River-derived water (swept channel + inflow) is 32 Mm³ in the same run, three-quarters of the total.

Melt reaches the official figure at exactly one point in the whole parameter space, scenario 8: 20.7 Mm³, and only by taking the largest published volume — six times the top of our envelope — with four-fifths ice, a 4,000 m drop and a 0.70 heat partition simultaneously. Relax any single one of those and it falls away fast; scenario 4, identical but for a 0.50 partition, gives 14.8.

So the balance of terms flips relative to Chamoli under monsoon + short fall + long runout, and the size envelope makes the flip decisive rather than marginal. The budget bounds the sources; only routing can test the arrival clocks and the Devghat peak. On to Phase B.

2. Channel profile & the snowplow-integral model¶

The Phase B model (model/snowplow.py) routes a front down the real channel. The picture: the debris/flood front advances at a slope-dependent speed; as it overruns the monsoon channel it incorporates ("integrates") the standing river water, collects lateral baseflow, liberates pore water from entrained saturated sediment, and carries the frictional meltwater generated in the steep upper reach. The wave body travels slower than the front, so the pulse stretches; the peak travels slower still and attenuates.

First, the channel itself: 499 points along the OSM-stitched river path (scar → Lhende Khola → Bhote Koshi → Trishuli → Devghat), elevations from Mapzen terrain tiles. DEM samples in a gorge sometimes catch canyon walls, so we enforce monotone descent and lightly smooth before taking slopes.

Front speed: fitting U = a·S^b + c0 to the observed clocks¶

The front-speed law has two parts: a slope-driven debris-flow/steep-torrent power term a·S^b, plus a flood-wave celerity floor c0 (deep-water √(gh) scale, ~8–10 m/s for h ~5–10 m) that carries the wave across the low-slope lower Trishuli, where observed celerity stays ~8–12 m/s even as S → 0.001. We fit (a, b, c0) by brute-force grid search against six front clocks and three peak clocks, weighted by data quality (collapse 08:37:10 NPT seismic = t 0; border CCTV 08:44 is the hard anchor). The peak travels slower than the front by a fitted factor c_p.

The fitted front hits the hard border clock (8 vs 7 min) and lands the Devghat front arrival exactly (403 min = 15:20). The speed panel shows the shape the fit had to find: ~50 m/s in the upper gorge collapsing to ~11 m/s at the Syabrubesi valley opening, then a slow decay to the 6–8 m/s celerity floor in the lower river — consistent with the geopera superelevation estimates and the station-to-station celerities, none of which were fitted directly.

Water bookkeeping: the snowplow march¶

Now march the front downstream and bookkeep water by source per km:

• melt, released over the steep upper reach (exponential, ~10 km scale), totalling the Section 1 best-evidence energy cap (5.3 Mm³);
• swept channel water = f_sweep · A_channel per metre of advance;
• baseflow inflow = Q_base · (time the front spends in each cell);
• pore water, released over the erosional reach (km 5–70);
• losses (overbank storage, deposition, stranding) at a fractional rate per km, attributed pro-rata so the source shares stay meaningful.

Monsoon baseflow is anchored at gauged/inferred points (border ~150 m³/s, Betrawati ~500, and the big Budhi Gandaki and Marsyangdi confluences taking the Devghat reach to ~1,500 m³/s). A station's peak discharge comes from a stretching skewed pulse: duration grows from ~45 min at the border to ~3.3 h at Devghat (the FFD stage record shows arrival 15:20, peak 16:00, back near normal 18:30), with F_WIN = 55% of the active water riding in the main window.

Four scenarios: snowplow (best evidence); melt-only (sweep-up and pore water switched off — the literal "it was melted glacier ice" narrative); melt-maximal with weak sweep; and snowplow + brief impoundment (ICIMOD's momentary damming of the Lhende).

Snowplow verdict. The best-evidence scenario puts 5,913 m³/s at Devghat against the observed 5,850 (+1%), with peak timing 448 min vs 443 observed and the front arrival exact (403 min = 15:20). Melt-only delivers 3,383 m³/s — 42% below the observed peak — and only ~4.4 Mm³ of genuinely new water against FFD's ~20 Mm³; even melt-maximal-with-weak-sweep falls 21% short. The distal wave is ~78% river-derived (swept channel + inflow); melt is ~8%.

Note the gross-vs-net distinction the bookkeeping makes possible: the snowplow gross surge at Devghat (~65 Mm³) is mostly redistributed river water that nets toward zero over a day-scale gauge integration (the emptied channel refills from baseflow), while genuinely new water (melt + pore) is ~10 Mm³. FFD's 20 Mm³ sits between the two — consistent with a windowed excess over a few hours (their 14:10–18:00 Devghat window), not with a melt-only source.

3. The ladder network: pulse structure as an R(L)C transmission line¶

The snowplow model routes a front and bookkeeps volume, but says nothing about pulse shape. For that, model/ladder.py makes the equivalent-circuit idea literal. The governing physics is the Saint-Venant shallow-water equations; dropping the inertia (momentum) term gives the diffusive-wave approximation, which is mathematically identical to a nonlinear RC transmission line — the telegrapher's equations without L, the same PDE as charging a long cable:

continuity:  dS_i/dt = Q_in,i - Q_out,i + q_lateral,i     (KCL at a node)
"Ohm's law": Q_{i->i+1} = (1/n) A R^(2/3) sqrt(dη/dx)     (nonlinear R)

River element	Circuit element
water-surface head η = z + h	node voltage
discharge Q	current
Manning reach friction (head loss ∝ Q²)	nonlinear series resistor
flow momentum (Saint-Venant inertia term)	series inductance
node storage S = w·dx·h	shunt capacitor C = w·dx
side valley + junction backwater	shunt RC branch: charges as the surge passes, discharges into the tail
temporary debris dam (crest erodes when overtopped)	breakdown element (SCR / spark gap); a cascade makes surge trains
baseflow & tributaries	distributed current sources

v2 installs the inductor. The first build dropped inertia entirely — an RC-only line — and honestly under-sharpened fronts. The upgrade keeps the ladder picture but makes discharge a state variable with its own momentum equation: the local-inertia form of Saint-Venant (Bates et al. 2010, the scheme inside LISFLOOD-FP),

dQ/dt = -g·A·dη/dx − friction        (the series L: current carries momentum)

integrated semi-implicitly in the friction term (stable against friction stiffness), with a Froude cap Fr ≤ 2.0 bounding the supercritical gorge reaches — the scheme drops convective acceleration, so the cap stands in for the physics it omits. simulate(inertial=True) is the default; inertial=False preserves the old diffusive/RC-only solver so the two can be compared like-for-like on identical geometry and forcing.

Domain: border (path-km 22, where the event became a flood) → Devghat (km 199), 90 nodes, dx ≈ 2 km, geometry from the same river profile. Channel width from downstream hydraulic geometry (clipped for gorge reaches); Manning n from rough mountain torrent (0.055) to big monsoon river (0.025). Inflow: monsoon baseflow plus a triangular event pulse at the border node — 30 Mm³ over 45 min, arriving at 08:44, i.e. the water the snowplow model says the wave carried at km 22 plus what the upper reach kept feeding.

Eight side-valley RC branches sit at the mapped confluences (Chilime, Langtang Khola, Mailung, Salankhu, Tadi, Mahesh, Budhi Gandaki, Marsyangdi), each a reservoir behind a broad-crested weir. Plan areas are order-of-magnitude reads of valley-floor storage — illustrative v0 values, to be mapped properly from Sentinel-2. The breach element is a dam crest that erodes when overtopped.

Four runs: bare line (inertial) / + side valleys (inertial) / the same side-valley configuration RC-only for comparison / + a hypothetical mid-route breach — labeled hypothetical because there is no evidence for one at km 55; it demonstrates the signature such a breach would leave in downstream records.

The Galchhi rate-of-rise is a genuine out-of-sample test — the "~9 m in ~30 min" observation was never used to calibrate anything. The RC-only v0 gave 5.2 m; installing the inductor lifts it to 5.9 m (and to 6.7 m when the ladder is re-run at 1-km node spacing) against ~9 m observed — movement in exactly the direction the circuit analogy predicts, because inertia is what lets a front steepen into a bore. The residual gap is honest physics and honest averaging: the local-inertia scheme still omits convective acceleration and hyperconcentrated-flow rheology, and each model node averages stage over a ~2 km reach and the full channel width, where the physical gauge sits at one constricted cross-section. Timing, by contrast, calibrates well (next cell). Pulse sharpness remains where the neglected physics lives — but less of it is neglected now.

What the inductor does, station by station (run 2 configuration, side valleys on; "excess" = peak above the local monsoon baseflow):

Station	RLC (inertial)	RC-only (diffusive)	effect of the inductor
Betrawati (km 68)	12,353 m³/s @ 09:40	10,617 @ 09:43	+17% excess, 3 min earlier
Galchhi (km 108)	6,279 m³/s @ 10:53	5,519 @ 10:59	+16% excess, 6 min earlier
Malekhu (km 117)	5,581 m³/s @ 11:15	4,940 @ 11:23	+15% excess, 8 min earlier
Devghat (km 199)	2,795 m³/s @ 15:28	2,707 @ 15:43	+3%, 15 min earlier

(observed: Devghat 5,850 m³/s @ 16:00)

Mid-reach peaks sharpen 14–17% and fronts arrive minutes earlier and steeper — the inductor does what an inductor should. The distal Devghat peak, however, barely moves. That persistence is the interesting result, taken up below.

Ladder verdict. Timing calibrates well against the clocks: inertial peaks at Galchhi 10:53 (observed rise ~11:00–12:00), Malekhu 11:15 (danger crossed 11:40 per the FFD release; 11:26 in early reports), Devghat 15:28 (observed 16:00). All three pulse-shaping signatures appear: the bare line attenuates and broadens; side-valley capacitors charge on the rising limb (clipping mid-route peaks) and discharge into the tail; the hypothetical breach stores the rising limb behind an eroding crest and lets go at ~2.8 h, re-sharpening the pulse — a double-peak visible at the mid-stations and smeared out by Devghat. That last signature is what to look for in the Gandak (Valmikinagar) record if a mid-route blockage is ever claimed.

And the honest accounting on the distal peak. The inductor sharpened mid-reach peaks 14–17% and improved the out-of-sample Galchhi test (5.2 → 5.9 m), but Devghat barely moved (2,795 vs 2,707 m³/s), and resolution and width sensitivities keep it in a ~2,500–2,900 band — the remaining gap to the headline 5,850 is not a missing term the inductor could supply. Part of the answer is likely the observation, not the model. FFD's own numbers do the arithmetic for us: ~20 Mm³ of excess water through Devghat in a ~3.8 h window (14:10–18:00) is a mean excess of ~1,450 m³/s, hence a triangular-pulse excess peak of ~2,900 m³/s — which implies the gauge's concurrent base flow was ~2,950 m³/s. That is a Narayani-scale base (Devghat sits below the Kali Gandaki confluence; likely DHM station 450), not the ~1,500 m³/s Trishuli-only base both our models route. On an excess-peak basis the target becomes ~2,900 m³/s, not ~4,350: the ladder (~1,300 excess) sits low and the snowplow (~4,400 excess) sits high, with the truth between the models. Confirming the Devghat station's river, rating curve and late-August base flow with DHM is now a top data task.

The division of labour stands: the snowplow model, which carries the front explicitly, is the volumetric workhorse; the ladder is the pulse-structure instrument.

---

4. Entrainment, and the falsification of the single-phase model¶

Built 3–4 September, three days after the sections above. This is finding 05 on the project home page, and until 5 September it was not in this notebook at all — the workings were published while lagging the findings by three days, which was wrong and is the reason this section exists.

The bug. Sections 1–3 model a flood that can put sediment down and never pick any up. A flood that visibly scoured a valley floor was, in our equations, incapable of scouring anything. Three observations said so, two of them quantitative, so the term could be built against literature constants and scored rather than fitted:

Target	Value	Source
Langtang corridor erosion	3.2 Mm³ eroded vs 0.9 Mm³ deposited over the ~45% mapped — net erosional ~3.5:1	geopera WorldView-3 stereo DEM, 1 Sept
Seti 2012 flow density	1.88 g/cm³ ⇒ w ≈ 0.47	SANDRP compilation, measured

Two closures, both with constants taken from the literature and nothing chosen with reference to those numbers. law="takahashi" is capacity-limited: an equilibrium sediment concentration for a mature debris flow on a bed of internal friction angle φ, eroding where the flow is under capacity and depositing where it is over. law="shear" is shear-driven (Frank et al. 2015), not capacity-limited. Full derivation in model/ENTRAINMENT.md.

The three cells below run the real model — model/unified.py, the same code the report uses, not a re-implementation. Each simulate() takes about 20 seconds.

The erosion volume lands, and the model breaks. 3.8 Mm³ of scour against 3.2 Mm³ measured, in the reaches they mapped, with nothing tuned — that is the term working. But with deposition now operating properly the same run buries 18 Mm³ in a corridor that can account for at most 12, and the front never reaches the border at all: inf means the flood stalled before kilometre 22.

That is not a parameter being slightly off. Before concluding anything, the honest move is to ask whether any combination of the contested inputs can satisfy both constraints at once. Rejection sampling over five inputs said no:

0 of 150 samples satisfied the upper-corridor clocks and the corridor
mass balance simultaneously. Border arrival and deposition were satisfied
together by ZERO runs; every other pair by 16-41.

Zero is a structural result, not a tuning failure. It falsifies the model form — the assumption that "sediment" is one substance with one settling behaviour — rather than the numbers in it. The fix separates the load in two:

• Coarse obeys the capacity closure, strands, and carries the granular friction. This is what builds a DEM-visible debris fan.
• Fine — silt and ice — is advected with the water and is rheologically part of the fluid (the friction dial sees w_eff = (h_w + h_f)/h). It does not deposit in this domain: wash load stays in suspension for hundreds of kilometres, and ice melts.

With f_fine = 0 the published tables of Sections 1–3 reproduce bit-identically, so this is an extension rather than a rewrite. The next cell runs both.

Read the second row against the first. Splitting the solids changes the erosion barely at all (3.80 → 3.75 Mm³, still against 3.2 measured) while the deposition falls from 18 to 5.4 Mm³ and the front arrives at 6.5 minutes against an observed 7.68. One release can now be dense enough to run fast and clean enough not to bury the valley, which is precisely the combination the single-phase sampling proved impossible.

The declared degree of freedom. f_fine is fitted, not derived. It is one number, it is declared here and on every summary page, and Section 5 replaces the hand-fitted 0.95 with a sampled posterior. The check that it is physics rather than curve-fitting is that two downstream measurements — Galchhi and Devghat — were held out of the fitting entirely; see Section 5.

Independent corroboration, weighted honestly. Darcy Weedman, writing at geopera, reached the same conclusion the same week by gradient-descent calibration of a 2D morphodynamic model against stereo imagery, which drove the settling velocity down by a factor of three and told him the material "behaves finer than assumed, remaining in suspension longer". Two unreviewed analyses agreeing is corroboration, not replication. geopera is Geopera Pty Ltd, a company blog, not an institution, and this notebook is not peer-reviewed either.

---

5. How big was it? The consistency envelope¶

Finding 04. Also built after Sections 1–3 and also missing from this notebook until 5 September.

Every public number for the release volume is a point claim from a different method, and they span a factor of four hundred: Kargel 50–200 Mm³, Azam/ICIMOD 100–200, geopera 60–140, the EGU hydrology blog 0.5–10. They cannot all be right, and arguing them one at a time has got nowhere.

So ask a different question. Not which number is correct but which combinations of inputs are simultaneously consistent with everything we can actually measure — which is a search over parameter space rather than a debate. This is approximate Bayesian computation in its simplest form: sample the contested inputs from stated priors, run the full model, keep the samples that reproduce every observable inside its stated tolerance.

Priors (deliberately wide, from research/event-dossier.md):

Input	Prior	Note
V_rel	log-uniform 1–200 Mm³	spans every published estimate
w0	uniform 0.02–0.95	liquid fraction; widened twice after the passing runs pinned against the ceiling
mu_dry	uniform 0.10–0.35	Schneider ice-avalanche to Scheidegger rock
n_scale	uniform 0.70–1.40	Manning roughness class-value uncertainty
h_erode	uniform 1–10 m	erodible layer depth
f_fine	uniform 0.0–0.98	the Section-4 split, sampled rather than hand-fitted

Observables — the upper corridor, where the hard data is:

Observable	Target	Tolerance
border arrival, km 22	7.68 min	±30%
Syabrubesi arrival, km 37.6	13 min	±50%
peak speed near km 22	48.5 m/s	±35%
erosion km 0–68	3.2 Mm³	±60%
bulk deposition	≤ 12 Mm³	inequality, not a target

Two things to flag before the result. The deposition constraint is an inequality: a DEM difference bounds deposition from above, it does not measure it. And the border-speed observable is the 45–52 m/s superelevation estimate; the first peer-reviewed study of the event (CAS, 1 September) measures 19 m/s at the same location, which is outside this tolerance. That re-scoring has not been done, and it is the largest single exposure in this section.

Regenerating the samples takes about 75 minutes:

python calcs/ensemble.py 220

so this notebook loads the saved output rather than re-running it. The array is committed to the repository, and the code that produced it is calcs/ensemble.py — read it alongside this.

The result, and its shape. 26 of 220 samples satisfy all five observables, and they put the release at 14–34 Mm³, median 21 — an order of magnitude below the largest published figures, and far narrower than the range we started from. This is an envelope from consistency, not a measurement of the mountain: nobody has surveyed the scar, and if someone does, this number is the one that should yield.

Only the volume is constrained. Liquid fraction, mu_dry and f_fine each span more than 90% of their priors among the passing runs — the right panel above. An earlier run appeared to pin all three, which turned out to be an artefact of a deposition cap of 5 Mm³ that was too tight; relaxing it to the honest 12 Mm³ released them, and those earlier tighter values are retracted. So we can say roughly how big, and we cannot say what it was made of, except that Section 3's clock excludes dry rock at any size.

Sensitivity to the shakiest input. The deposition cap is the weakest number in the scoring. Relaxing it by 2.4× moved the envelope about 30% and the two ranges overlap, so the answer is being driven by the measurements rather than by that constraint.

The out-of-sample half-failure. Galchhi and Devghat were held out of the scoring entirely so they could act as an honest check on the passing runs:

• Devghat peak discharge: 1,752–2,889 m³/s against ~2,900 observed. All 26 meet the stated factor-of-two criterion — but all 26 also land below the observation, so the correct reading is "not contradicted", not "reproduced", and the residual is one-sided.
• Galchhi 30-minute stage rise: 1.8–6.1 m against ~9 observed. 23% pass. The model routes about the right discharge and converts it to the wrong stage, which implicates lower-reach channel geometry — below km 60 our widths come from hydraulic geometry, not measurement.

We publish the half that failed rather than the half that worked. The full Galchhi hydrograph from DHM is the single dataset that would most sharpen this.

Findings, limitations, pointers¶

Findings (all as envelopes over the contested inputs — source volume 10–200 Mm³, ice fraction unpublished):

• The distal flood wave was overwhelmingly river-derived. In the best-evidence snowplow scenario, ~78% of the active water at Devghat is swept channel water + baseflow collected en route; frictional ice melt is ~8% (~3.7 of 47 Mm³) in that v1 100 Mm³ run; at the 14–34 Mm³ envelope of Section 5 it is 1.1–2.5 Mm³, a few per cent (scenarios 6–7 above).
• Melt-only fails. Switching off sweep-up leaves 3,383 m³/s at Devghat vs 5,850 observed — 42% low — and ~4 Mm³ of new water vs FFD's ~20 Mm³ excess. No defensible input set rescues it: melt is energy-limited (~12 kJ/kg from the initial fall against 334 kJ/kg latent heat), and even the melt-maximal steel-man budget at the envelope caps at 2.5 Mm³; only the 200 Mm³ every-dial case (scenario 8) reaches 20.7.
• The snowplow scenario matches the observations it was not fitted to: Devghat peak 5,913 vs 5,850 m³/s, front arrival 403 min (15:20) exact, peak 448 vs 443 min, gross/net water bracketing FFD's windowed 20 Mm³.
• Chamoli 2021 was the exception, not the template: the same arithmetic reproduces its ~5 Mm³ ice-limited melt (~87% of the water), and shows why the balance flips under monsoon flow + short fall + long runout.
• Pulse structure is circuit-like: the ladder network reproduces the station clocks (Galchhi 10:53, Malekhu 11:15, Devghat 15:28), demonstrates side-valley charge/discharge, and gives the breach double-pulse signature to seek in the Gandak records. Installing the inductor (local-inertia Saint-Venant) sharpened mid-reach peaks 14–17% and moved the out-of-sample Galchhi 30-min rise from 5.2 to 5.9 m (vs ~9 m observed) — the RC→RLC upgrade acts exactly as the circuit analogy predicts.
• The Devghat "observed peak" needs unpacking: FFD's own volume/duration arithmetic (20 Mm³ over ~3.8 h → excess peak ~2,900 m³/s) implies the gauge's base flow was ~2,950 m³/s — Narayani below the Kali Gandaki confluence, not the ~1,500 m³/s Trishuli-only base the models route. On an excess basis the ladder sits low and the snowplow high, with the truth between them.

Limitations. Source volume and ice fraction are unpublished — everything above is an envelope, and FFD's 20 Mm³ is single-source with unpublished method. The front-speed law and loss rates are calibrated, not derived. Side-valley plan areas are order-of-magnitude v0 values pending Sentinel-2 mapping. Even with the local-inertia term the ladder under-predicts the distal peak against the 5,850 headline (though much of that gap is likely the gauge's ~2,950 m³/s Narayani base flow rather than missing physics — confirming the Devghat station with DHM is the top data task) and still under-sharpens the Galchhi rise ~35%. The DEM profile is Mapzen terrain in a gorge — monotone enforcement and smoothing are load-bearing. geopera velocity/height figures are provisional.

Also unresolved, added 5 September. The border speed that Section 5 scores against (48.5 ± 35% m/s) is contradicted by the first peer-reviewed study of the event, a CAS frame-by-frame video analysis measuring 19 m/s at the same place. The envelope has not been re-scored against it. And the modelled debris deposit sits at km 0–36 where stereo measurement puts it at 40–43 — geopera's model gets that right and ours does not.

Pointers. All links go to the live site or the repository, so they work from wherever you are reading this.

• Write-ups: technical report · plain English · project home
• Plan and honesty rails: PLAN.md
• Models: calcs/energy_water_budget.py · model/snowplow.py · model/ladder.py · model/unified.py · model/core.py · calcs/ensemble.py (this notebook reproduces their published tables; the ladder's v3 junction step and the 75-minute ensemble are loaded from saved output rather than re-run here)
• Entrainment build: model/ENTRAINMENT.md
• Profile construction: model/build_profile.py · model/fetch_elevations.py
• Evidence base: research/event-dossier.md · research/science-review.md
• Corrections: open an issue — they are recorded on the project home page with the date and the reason.

Dave Hume, with Claude as research and modelling assistant. Sections 1–3 2 September 2026; Sections 4–5 added 5 September 2026.