The flood started in Rasuwa on Wednesday morning. Rasuwa has recovered thirteen bodies.

Chitwan district, roughly two hundred kilometres down the Trishuli and never touched by the glacier that began this, has recovered 259 — more than a third of every body Nepal has found. Nawalparasi East, further down again, has 184. Nawalparasi West has 82. Three lowland districts that experienced this disaster as a rise in the water level are holding 525 of the country's 734 dead.

Rasuwa is holding 562 of the missing.

That is the shape of the Bhotekoshi disaster four days in. The deaths happened in the mountains and the counting is happening in the plains, and between the two lies several hundred kilometres of river that is still delivering.

Nepal's National Disaster Risk Reduction and Management Authority put the figures at 9 a.m. Sunday: 734 bodies recovered, 2,498 people unaccounted for, 242 injured. Across the border in Gyirong County, Chinese authorities reported 16 dead and 546 missing, of whom 261 were foreign nationals. Combined, that is 750 confirmed dead and 3,044 people the two governments cannot locate. The United States Geological Survey attributes the event to a glacial collapse on the Tibet–Nepal border that sent a wall of ice, rock and mud down the Bhotekoshi and into the Trishuli.

The more revealing document is the missing list, because it is not a census. It is a set of manifests. Nepal's authority counts 933 people linked to hydropower projects on the affected rivers. It counts 589 foreign nationals — trekkers, and pilgrims heading to Mount Kailash — plus 127 Nepalis travelling with them. It counts 85 security personnel: 45 from the Nepali Army, 27 from Nepal Police, 13 from the Armed Police Force. Add those categories together and you have most of the 2,498. Almost nobody on that list was in the gorge by accident. They were there because a job, a permit, a pilgrimage route or a duty roster put them there, on rivers that Nepal has spent a decade fitting with turbines and has not yet fitted with an upstream gauge.

The rescue has narrowed to one of those manifests. At the Trishuli 3A hydroelectric site in Rasuwa, army engineers spent Saturday excavating a tunnel mouth with drilling equipment — 82 army and police personnel on the entrance alone — and by Sunday had pushed a pipe into the passage where more than a hundred workers are believed to be. The army says it pulled 350 people out of the site on Thursday under conditions the prime minister's office called extremely risky. India has sent a specialist tunnel rescue team, which began with an aerial survey. China has moved more than 2,100 rescuers to the Gyirong side, allocated 220 million yuan, about $32.7 million, and offered satellite and hydrological data along with tunnel experts.

The hydrology is not finished. Drones detected a landslide about 2.8 kilometres downstream of the lake left by the glacier collapse, forming a second dam and a second pool, according to Chinese state broadcaster CCTV; Chinese rescuers at the worst-hit site were moved to higher ground on Sunday. Nepal's disaster authority issued its own warning the same morning that flow in the Bhotekoshi at Rasuwa had increased again. Fresh rain lifted the Trishuli. Police told Reuters that residents and rescue workers moved uphill.

Which leaves the part nobody has solved. Bodies pulled from the Narayani in Chitwan are not identifiable by the people who find them. Nepal has distributed unidentified remains to 21 hospitals across multiple districts for preservation and identification, and 589 of the missing hold foreign passports, which turns identification into a consular and forensic problem running in parallel with the rescue. Nepal, China, India and South Korea have all put technical or forensic personnel into the field.

One discrepancy we cannot reconcile and will not paper over: Sunday's Nepali government update lists 7,514 people rescued, while wire reporting on Saturday and Sunday continued to use a figure above 3,700. We are printing both numbers rather than choosing the flattering one.

The figure to watch this week is not the death toll. It is the ratio. Three thousand and forty-four people are missing against 750 confirmed dead, and the water arrived four days ago. The dead are not going to be found where they were lost. They are going to be found in Chitwan, and in Nawalparasi, and further down, by people who were never in any danger, over a period that will be measured in weeks.

At about 8:10 p.m. Friday, a Russian Geran-4 jet-powered drone struck a facility in Myla, a village five kilometres west of Kyiv on the Kyiv–Chop highway. The facility was a Ukrainian military ammunition depot. Shells, mines and drone components began detonating and kept detonating. Fire moved through the village.

Thirty-seven people were killed, the deadliest single attack of the war this year, surpassing the 31 killed in Kyiv on 2 July. Forty-two were injured, four of them children. Between 50 and 130 residential buildings were damaged, depending on which official count you take, and 381 people including 59 children were evacuated.

Thirty-four of the dead were residents of a care home for elderly and disabled people, according to Tymur Tkachenko, who heads the Kyiv regional military administration. They lived next to the depot. Kyiv Oblast has declared Sunday a day of mourning; flags are at half-mast and entertainment venues are restricted.

Ukraine's prosecutors opened two investigations. One is the expected one, under the criminal code provisions covering Russia's crime of aggression. The other is directed inward: whether the storage of explosive objects and materials on those grounds was lawful, and whether the company responsible held the necessary permits and approvals. The company has not been named.

What makes this more than a wartime tragedy is that it is a repeat. In July, a munitions depot at Vyshneve, also near Kyiv, was hit in a deadly attack, and the resulting investigation found serious failings by the state-owned operator, which had not complied with standards despite the proximity of residential buildings. Prime Minister Sergiy Koretsky said as much on Saturday. "In the fifth year of the full-scale invasion, making the same mistakes over and over is criminal negligence," he wrote, adding that everyone responsible must face severe punishment.

That is a head of government describing his own state's conduct in the language of the criminal code, on the day of a massacre carried out by an invading army. It is worth sitting with how unusual that is, and how correct.

The rest of the night ran as these nights run. Russia launched at least two missiles and 267 drones between Friday evening and Saturday morning, killing three more people including a two-year-old in Kyiv and a fourteen-year-old girl in the suburbs. Russia's defence ministry said it hit a warehouse in Myla storing components and launch boosters for long-range drones, and also struck the ports of Mykolaiv and Izmail. Moscow said it destroyed 313 Ukrainian drones overnight. Ukrainian strikes wounded eleven people in Rostov region, killed three in Belgorod, and killed one in Sevastopol. Zelensky told the country Friday that Ukraine intends to scale its long-range drone launches to a thousand a day.

Both countries are now running industrial-scale munition flows through logistics networks embedded in towns. Myla is what that looks like when one of the nodes is five kilometres from a capital and next door to a building full of people who could not run.

Begin with what has actually been released: a Truth Social post, a statement from Caracas, and a White House background description. That is the documentary basis for what the President of the United States called "THE BIGGEST OIL DEAL IN WORLD HISTORY." No text of the agreement has been published.

The reported terms: the US government and an unnamed private Venezuelan operator have formed a new company holding rights to 17 fields with 65 billion barrels of proven potential, for one hundred years. Washington takes what has been described as 55 per cent of effective output, comprising an ownership stake and the right to buy oil at cost. Interim President Delcy Rodríguez says the arrangement draws $100 billion of investment and yields Venezuela more than $209 billion in taxes.

Now the number that has not been discussed. Rodríguez put that $209 billion figure at a benchmark of $65 a barrel, and said explicitly that prices could fluctuate. Brent closed the week near $88, down more than five per cent over five sessions and still carrying something like a 25 to 40 per cent premium over pre-crisis levels, six months into a Strait of Hormuz blockade that has not been lifted.

So Caracas has published a headline revenue estimate that requires crude to fall roughly a quarter from where it trades today. Either the government is being conservative in a way that governments announcing historic deals are almost never conservative, or the number was built on an assumption that the war ends and the Hormuz premium goes with it. There is no third reading. Anyone forecasting Venezuelan fiscal revenue off the $209 billion should say which one they believe.

The timing question is separate and simpler: nothing changes at the pump. The national average is $4.08 a gallon, against $3.20 a year ago. Patrick De Haan of GasBuddy told TIME that drilling and pumping this oil "will take a very long time," and that changes to fuel prices will not happen overnight or in months, with global refining capacity constrained on top of that. The fastest route to lower American gasoline prices runs through the Gulf, not the Orinoco. The deal covers reserves in the ground, not barrels on a ship.

What is real and immediate is the trade that already exists. The United States has more than quadrupled crude imports from Venezuela this year, to roughly 600,000 barrels a day, making Venezuela the second-largest source of imported oil after Canada, per EIA data. Venezuelan output in July ran about 1.16 million barrels a day, half what it was a decade ago. Friday's announcement mostly formalises and extends a flow that resumed after American forces removed Nicolás Maduro in January.

One further item worth watching before OPEC+ convenes on 6 September: energy trade reporting has Caracas weighing an exit from OPEC, discussed alongside the American access talks, with no decision taken. At 1.16 million barrels a day, Venezuela's departure would be arithmetically trivial and institutionally loud.

Our position for the week: we will not model a contract we have not read. When the text is published, we will price it. Until then, the only hard number in this story is $88, and it is going the wrong way for Caracas's forecast.

The document released Friday by American Oversight, obtained through a Freedom of Information Act request to Immigration and Customs Enforcement, is an April 2025 memo from the Department of Homeland Security's Office of the General Counsel. Politico reported it first; NOTUS also received it.

The legal construction in it has two steps. First, federalize National Guard members under Title 10. Second, designate them as immigration officers under 8 U.S.C. § 1103(a)(6). DHS's position is that because Congress authorized that designation, troops so deputized can perform civilian immigration enforcement — including arrests of unauthorized immigrants, including outside their home states — without running into the Posse Comitatus Act, the 1878 statute restricting military policing inside the United States.

Two things about that memo are worth separating.

The first is the legal argument, which scholars quoted this week dispute on the straightforward ground that Congress did not write immigration law as a general exemption from Posse Comitatus. The administration has already lost ground here. A court rebuked the Los Angeles deployment for running afoul of Posse Comitatus. In December the Supreme Court held that local forces must be incapacitated in some way before military personnel can be used for domestic purposes, a ruling that followed Illinois's challenge to Operation Midway Blitz. Guard troops have since done background immigration work and provided security for federal agents making arrests; the administration has not openly ordered service members to make the arrests themselves.

The second is who received it. The memo went to all state attorneys general. It was not a private opinion filed in a drawer at DHS; it was a notice to the officials most likely to sue, sixteen months before it became public. Whatever else it is, it is a position the department was willing to state to its adversaries.

Which is why the timing of its surfacing matters more than its age. The theory as drafted is not limited to a particular set of circumstances. American Oversight's executive director, Chioma Chukwu, argued that combined with the administration's escalating and unsubstantiated claims about noncitizen voting, it supplies a ready-made rationale for placing federalized troops at or near polling places under an immigration-enforcement banner, ten weeks before a midterm election. Guard troops have been patrolling near the White House as recently as 17 August.

That is a forecast, not a finding, and we mark it as one. But the mechanism it describes is not speculative. It is written down, it was mailed to fifty state attorneys general, and nobody has withdrawn it.

OpenAI said Friday it will stop providing its models to Cursor, the coding tool now owned by Elon Musk's SpaceX. The stated reason, from OpenAI's own blog post: "We cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts."

Read that sentence for what it withholds. It does not allege that Cursor violated anything. It cites experience with other companies under the same owner and extends the inference to this one. The unit of trust is not the product or the contract. It is the parent.

Musk's response, posted early Saturday: "I couldn't care less." Michael Truell, a Cursor co-founder now an executive at SpaceX, said the company was talking to the OpenAI team. Anthropic said it will increase compute support for Claude models inside Cursor, which is the commercially obvious move and also a quiet advertisement: we do not do this.

The interesting question is not whether the feud is petty. It is what the episode establishes about the shape of this industry. For three years the pitch to founders has been that frontier models are infrastructure — that you build on an API the way you build on a cloud region or a payment rail, and the substrate is indifferent to you. Utilities do not read your shareholder register. AWS did not cut off customers over their acquirers.

Model providers, it turns out, do. The supply relationship is discretionary, the terms of service are elastic enough to cover an owner's reputation, and the switching cost is borne entirely downstream. Cursor is unusually well-positioned to absorb it: it has a rival provider eager to backfill, an owner with capital, and an in-house model effort. A smaller company in the same position has no such options, and every acquisition conversation in this sector now includes a new item of diligence — whether the buyer's name is acceptable to the sellers of the underlying models.

That is a real change in the market structure, and it happened in a blog post about a personal rivalry. Those are usually the ones that stick.

SPRINGFIELD, GEORGIA — About 75 people filled the county chambers on Saturday morning and another 25 listened from an overflow room across the hall. It rained. Protesters stood outside the administrative complex with signs. Inside, for nearly three hours, residents questioned commissioners about OpenAI's proposed $20 billion data centre at the Savannah Gateway Industrial Hub in Rincon — noise, water, electricity, emergency services, and how the whole thing was handled.

No one from OpenAI attended. Board Chairman Damon Rahn was also absent.

County Manager Tim Callanan opened with a history of the Industrial Development Authority, emphasising the IDA's independence from the elected board of commissioners. That is a legally accurate point and, in a room full of people who found out about a 3.2-gigawatt facility in July as a finished fact, an unhelpful one. It amounts to telling voters that the body that made the decision is not the body they elected, which is precisely the complaint.

The demand that came out of the meeting has hardened in a way worth noting. Residents are past wanting assurances. They want contractual obligations rather than promises, and information from neutral sources rather than from the company. Two commissioners, Burdette and District 4's Beth Helmly, acknowledged a moratorium on future data centres may be considered — Helmly framing it as time to write stronger ordinances and to make data centres a conditional use requiring additional review.

"Conditional use" is the phrase to watch. It is zoning language, it is boring, and it is the only mechanism in this story that would actually give a county leverage before a deal is signed rather than after.

The politics have scaled up. Senator Raphael Warnock, speaking Friday at the site, called for a statewide moratorium, with an exemption structure that reads as a list of grievances: projects could proceed if they do not raise water or electric bills, do not rely on non-disclosure agreements or sweetheart tax giveaways, and do not cause air, water, noise or light pollution. Georgia regulators approved a Georgia Power contract last week to supply 3,200 megawatts to the site. Georgia Power's Kim Greene has said the company's arrangements are structured so other customers do not subsidise the load; OpenAI says it will pay hundreds of millions in state and local taxes, expects to be the county's largest taxpayer, and will not raise residential electricity rates or divert community water.

Georgia now hosts 276 data centre projects, fourth in the country behind Virginia's 674, Texas's 537 and California's 296. In April, Decatur County's commissioners imposed a local moratorium on permits and plan reviews for data centres — and, in the same measure, motocross complexes — while a committee studied the question. That pairing tells you how the category is currently understood in rural Georgia: as a land use with noise, not as an industry.

Saturday's session was the third public meeting in Effingham on data centres in a week. Commissioner Burdette suggested the next one should include the company. That would be the first.

The facts, in the order that makes them hurt.

Warner Bros. Discovery finished a Looney Tunes feature and decided it was worth more shelved than shown, because a completed unreleased film can be written off and a released flop cannot. Ketchup Entertainment, an indie distributor, bought it out of the vault for $50 million and spent roughly $10 million marketing it. This weekend it opened in 3,573 theatres, took $5.2 million on Friday, and is tracking to $13–14 million for the frame.

That is the largest opening in Ketchup's history by a factor of more than four; the previous record was $3.1 million, set by another Looney Tunes rescue, *The Day the Earth Blew Up*, in March 2025. Critics are at 94 per cent, audiences at 97. It earned an A CinemaScore — the first Looney Tunes theatrical feature ever to do so, and a grade only two other Warner-produced animated films have matched: *The Iron Giant* in 1998 and *The Lego Movie* in 2014.

Those two comparisons are not flattering to the studio. *The Iron Giant* was botched on release and became canonical anyway. *The Lego Movie* was a franchise the company then ran into the ground. Warner Bros. has a long, consistent record of being the last party in the room to understand what it owns.

Now the sober part, because triumphalism here is cheap. Fourteen million against a $60 million all-in cost is not vindication in the accounting sense; Ketchup will need long legs and a very good home window. Awareness was the film's problem all weekend, and a $10 million marketing spend does not buy awareness for a title whose primary fame is the story of its own near-destruction. The write-off math may still turn out to have been correct as math.

That is the actual scandal, and it is not about one film. The tax code has become a critic with a vote, and its aesthetic criterion is that an unseen object is more valuable than a seen one. Under that rule, the correct move for a studio holding a good film with a soft commercial profile is to burn it. *Coyote vs. Acme* only survived because its destruction became embarrassing enough to be a story.

Meanwhile, at the top of the chart, the machine that never has this problem: *Spider-Man: Brand New Day* took its fifth number-one weekend with about $21–22 million, ran to roughly $866.5 million domestic after 27 days, and passed *Titanic* worldwide at $2.32 billion to sit fourth all-time behind two *Avatars* and *Endgame*. It is about $70 million from taking *The Force Awakens*' domestic record. Nobody was ever going to shelve that.

And Ridley Scott's *The Dog Stars* opened around $7–8 million with a 39 per cent critics' score and a C+ from audiences, which is what the market for an actually unwanted film looks like. Note the difference. A studio can tell the two apart. It is simply not paid to.

Go see the coyote. Not as a protest — the movie is good. But it is worth remembering that you were not supposed to be able to.

The result: 99,037 against, 89,459 for. 52.54 to 47.46, on 69.56 per cent turnout, out of 188,496 valid votes. Iceland will not reopen accession negotiations with the European Union. The referendum was technically non-binding; Prime Minister Kristrún Frostadóttir had already promised to respect it, and by her own framing a No puts the question away for the rest of a term running to 2028.

The post-mortems will say fish, and they will be partly right. Iceland's exclusive control of its waters is not a bargaining chip in Reykjavík; it is the thing the country is. Brussels was reportedly willing to be flexible on it, which is precisely the offer that fails, because a flexibility extended is a flexibility that can be withdrawn, and everyone in a fishing town understands the difference between a right and an accommodation.

But the fish argument was available in 2009, and in 2013 when a Eurosceptic government killed the talks, and it did not require this referendum to be settled. Something else was on the ballot.

Consider how the case for Yes was constructed. Iceland was described by European officials as a low-risk accession that could soften scepticism toward enlargement and strengthen the bloc at a moment when Arctic security is in focus. Read that from a kitchen in Akureyri. It says: your membership would be useful to us. It says: you are small, uncomplicated and well-behaved, and your accession would help our other conversations. It arrives while an American president talks openly about acquiring Greenland, which is to say while the North Atlantic is being reorganised by larger powers discussing the disposal of small ones.

The government compounded it by calling this a now-or-never moment. That is a closing technique. Applied to a nation deciding whether to hand a supranational body authority over its principal industry, it reads as pressure, and pressure applied to Icelanders has a reliable historical yield.

So the vote was not really about Europe, which most Icelanders like, and not only about cod. It was about the difference between joining an institution and being recruited into a demonstration. Brussels wanted a proof of concept. Voters were told that their yes would help ease scepticism elsewhere, strengthen a bloc's strategic position, and reassure officials about enlargement — all real arguments, none of them about Iceland.

Here is the part Brussels should find genuinely uncomfortable: this was the easy one. Wealthy, stable, already inside Schengen and the single market through the EEA, no rule-of-law dispute, no border problem, no oligarchs. If the enlargement pitch cannot clear fifty per cent in Iceland in a year when the Arctic has never mattered more, the pitch is the problem. The bloc has spent two years arguing that geopolitics makes membership urgent. Urgency is a good argument for a chancellery and a poor one for an electorate, because a country told it must join for everyone's sake immediately hears that it is being asked for a favour, and countries that are asked for favours want to know what they get.

Iceland's answer, by five percentage points: not enough, and not now, and stop telling us it is now or never.

There is a version of Friday night that is complete and true and requires no further reporting. A Russian drone struck a village five kilometres west of Kyiv. Thirty-seven people died. Thirty-four of them were elderly and disabled residents of a care home. Russia did that. Nothing Ukraine did or failed to do transfers an ounce of that guilt.

We published a second set of facts alongside it: that the site was a military ammunition depot, that Ukrainian prosecutors are investigating whether the explosives were stored there lawfully and whether the operator held permits, that a nearly identical detonation killed people at Vyshneve in July, and that the earlier investigation found the state-owned operator had not complied with standards despite the proximity of homes.

We expect letters. The argument against us is not stupid, and I want to state it in its strongest form rather than a convenient one. It goes: attribution is the whole story in an aggressive war; every column inch spent on the victim's paperwork is an inch Moscow's information apparatus will clip and repost; a country under nightly bombardment cannot run a peacetime siting process; and a newspaper that treats both facts as equally newsworthy has quietly imported a false symmetry into a war that has none.

Three responses.

First, we did not have to go looking. Ukraine's own prosecutor general opened the negligence investigation. Ukraine's own prime minister, Sergiy Koretsky, wrote the words: in the fifth year of the full-scale invasion, making the same mistakes over and over is criminal negligence. A foreign paper suppressing a self-criticism that Kyiv published itself is not protecting Ukraine. It is deciding that Ukrainians can handle a fact that our readers cannot.

Second, the practical stake is not reputational, it is thirty-four people. The gap between Vyshneve in July and Myla in August is seven weeks. Whatever review followed the first one did not move the ammunition away from the second care home. That is not an abstraction to be weighed against morale; it is a specific, correctable failure with a body count attached, and the only mechanism that corrects it is pressure. Silence in the name of solidarity is a decision to let the third one happen.

Third, on symmetry: we are not confused about who fired. Russia launched two missiles and 267 drones that night and killed a two-year-old in Kyiv and a fourteen-year-old in the suburbs, and we said so. Reporting that a state also failed its own people is not the same as splitting the blame in half. Institutions can be attacked and negligent in one event. Most institutions in most disasters are.

Where would this rule break? I think there is a real line at operational detail — locations, capacities, dispersal patterns, anything that improves targeting rather than accountability — and we will not cross it. I am less certain about timing. Publishing on the day of a regional day of mourning is a choice, and a reasonable editor could argue for waiting a week and losing nothing but attention. I chose not to wait, because seven weeks is the interval that matters here and a week of ours is a meaningful fraction of it.

If you think that is wrong, write and say so, and be specific about which of the three responses fails. We will print the best objection we get, including if it is better than what I have written here.