The Filtration Gamble

By the afternoon, China's customs administration had suspended imports of all Japanese aquatic products. Not from Fukushima. From Japan.

The water is almost certainly fine. This is the sentence that gets said first and settles nothing. Tritium is diluted to under 1,500 becquerels per litre — one-fortieth of Japan's own regulatory ceiling of 60,000, about one-seventh of the World Health Organization's drinking-water guideline of 10,000. The annual tritium cap is 22 terabecquerels, which is the operational target the plant ran under before the accident, back when it was a normal power station doing normal things. In July the IAEA published a comprehensive report concluding that the plan is consistent with international safety standards and would have a negligible radiological impact on people and the environment.

None of that is in serious dispute among the regulators who reviewed it. It is not unanimous, and the dissent is worth a second of your time. The U.S. National Association of Marine Laboratories, which speaks for more than a hundred member labs, objected that the sampling protocols and statistical design were not adequate to support the safety claim. The Pacific Islands Forum's independent expert panel concluded the discharge had not been proven safe and should not proceed. Neither of those is a fringe position, and notice what they are actually contesting: not the physics of tritium, which is settled, but the adequacy of the data. Hold onto that, because it turns out to be the whole essay.

And on the day it took effect, the country next door banned every fish.

The comfortable read is that this is politics beating science — a sound engineering case losing to nationalism, panic, and a trade dispute in a lab coat. There's something to it. It also lets the losing side off a hook it built itself. The hook is the interesting part, because it isn't really about nuclear power.

i · the name shipped before the machine did

ALPS stands for Advanced Liquid Processing System. Its job is to strip 62 radionuclides out of the water that has been circulating through three melted reactor cores since 2011. It cannot remove tritium — tritium is hydrogen, it rides inside the water molecule itself, and separating it at this volume is not industrially realistic. That limitation has been stated openly from the start. Fine. Honest constraint, honestly disclosed.

The problem is the other 62.

In August 2018 — during the public hearing process, which is its own kind of detail — it emerged that ALPS had not been doing what TEPCO said it was doing. Reporting by Kyodo and others established that roughly 70 percent of the stored water still exceeded release standards for radionuclides other than tritium. Strontium-90 in some tanks ran to a hundred times the limit. In the worst tanks, the figure reached roughly twenty thousand times the limit. At the hearings themselves, the data presented had been the data that complied.

It would be too neat to call that the moment TEPCO's credibility died. Its credibility was in poor health well before. The company falsified inspection and repair records at its reactors in 2002, a scandal that cost it its chairman and its president — nine years before the accident. It took two months in 2011 to confirm that three cores had melted. It withheld the SPEEDI dispersion projections while people evacuated, in some cases directly into the plume. In 2013 a prime minister told the International Olympic Committee the situation was "under control" while contaminated groundwater was reaching the sea.

So 2018 is not where trust died. It is something narrower and, for this particular thirty-year plan, worse: it is where one word stopped working. Every load-bearing element of an ocean discharge leans on the adjective in "ALPS-treated water." That adjective is what the IAEA certifies. It is what the fishing cooperatives are asked to accept. It is what the dashboard will report every month until 2053. And in August 2018 the public learned that the adjective had been describing the machine's job title rather than the water inside the tank.

Give TEPCO this much: they fixed it. Secondary reprocessing puts the non-compliant tanks back through ALPS, the current batch has been sampled and independently verified, and the engineering has genuinely caught up to the name. That is real work, it took years, and it is not nothing.

It also has no announcement attached to it. There is a press conference for the launch of a purification system. There is no press conference for the day, five years later, when it finally does what the launch said it did. The distance between those two dates is where the credibility went, and everything since has been an attempt to buy it back at a fixed exchange rate in a market that doesn't offer one.

ii · a certificate is a snapshot. trust is everything that happened.

Here is the structural asymmetry, and it generalizes well past reactors.

Verification answers a question about a system's state at a moment: is this water, in this tank, today, inside these limits. That is what a regulator does, and the IAEA's report is scrupulous about its own scope — it assesses the discharge as planned, as engineered, and as sampled. It is not a character reference for TEPCO. It cannot be. That isn't a thing an inspection produces, and a regulator that started producing them would be worth less, not more.

Trust is a different quantity entirely. It's computed over the whole history: every prior claim, checked against what turned out to be true. It does not reset when the current reading comes back clean. It can't, or it would be useless — a measure that forgives everything before today is just a measurement.

So Japan spent two years running a campaign that could not work on the timescale it needed to. The strategy was: demonstrate the technology, publish the data, invite the inspectors, and public opinion will follow. Every one of those moves improves the snapshot. Not one of them touches the history.

You cannot fix an integral by improving the value at one endpoint. You can move it — that is the honest version, and the correction matters. Keep accumulating clean values and the average does drift upward. This is not a locked door. It is a very slow one. But slowness is the whole finding, because trust recovers at a rate no political actor can afford to wait for. Ministers have four-year horizons; the integral has a thirty-year one. "Publish verified data every month for three decades" is a correct strategy that nobody in the building will still be employed to see through. That is not the same thing as impossible, and it produces the same outcome for everyone currently alive and holding a fishing licence.

And there's a second breach, one that people register far more readily than they register becquerels. In August 2015, the government and TEPCO gave the fishing cooperatives a written undertaking: no disposal would proceed "without the understanding of interested parties." By this summer, ministers were describing the industry as having reached "a certain level of understanding" of the plan. In June, Masanobu Sakamoto, who heads the national fisheries federation, said it plainly: they could not support the government's position that an ocean release was the only solution.

He was not confused about the plan. He understood the plan. He objected to it.

Somewhere between 2015 and now, understanding was quietly re-scoped from consent to awareness. That is not a radiological event and no instrument on the site will detect it. It is also the kind of thing a fisherman can verify himself, without a spectrometer, in about four seconds. Compensation funds were announced ahead of the release; long-term support for the industry was pledged three days ago. Money is what you offer when the word you gave has been redefined and both parties know it.

iii · the ban is not about the water

China's suspension is not a safety measure and it is not especially subtle about being something else. Figures published by Japan's environment ministry put the Qinshan Phase III plant in Zhejiang at roughly 143 terabecquerels of tritium discharged in 2020 — something over six times Fukushima's annual cap, from a facility that never melted down. Chinese coastal plants routinely put more tritium into the same ocean than the plan Beijing has spent months describing as an assault on the marine environment.

That's worth saying. It is not a defense, and it is a narrower point than it looks. Qinshan's cooling water has never touched melted fuel; it needs no ALPS because it has no 62 other radionuclides to strip. Comparing the two on tritium alone is precisely the like-for-unlike move this essay has spent a section teaching you to distrust. What the comparison establishes is that Beijing's framing is opportunistic — not that the two discharges are equivalent. That is enough. It is also all it gets.

Because the hypocrisy charge only bites where your own record is clean, and Japan's isn't. Once the label and the contents have visibly come apart — once "treated" turned out to mean treated in the sense that we ran it through the thing named treatment — an opportunistic actor doesn't have to manufacture anything at all. The ammunition is in your own disclosure. You generated it, you published it, and now somebody else is using it for a purpose you would never have chosen. There is no counter-move available, because the counter-move is trust us, and that is the precise instrument that broke in 2018.

South Korea is the cleaner measurement, and it's cleaner precisely because Seoul did not fight it. The government reviewed the plan and said the science held. Its citizens went out and bought salt. Sea salt prices climbed roughly 27 percent from April; by July the government was releasing 50 tonnes a day onto the market at a 20 percent discount — a state simultaneously certifying that the sea is safe and subsidizing a panic about the sea. Both actions were sincere. Neither was irrational for the actor taking it.

That gap, rendered as a commodity price, is the cleanest reading anyone is going to get: a state that verified it and a public that didn't, sharing one country, one shoreline, one set of published numbers. The verification was real. It just wasn't the thing being consumed.

iv · nobody holds a press conference for the fix

Which brings us back to the smallest observation in this piece, the one that turns out to be the largest.

There was a launch for ALPS. There was no launch for the day, five years later, when ALPS started doing what the launch said it did.

That asymmetry is not a TEPCO failing. It is structural, and it is everywhere. Institutions have elaborate, well-funded, ceremonial machinery for making claims — ribbon-cuttings, press kits, embargoed briefings, a report release with a director general at a podium and a room full of cameras. They have essentially none for retiring a bad claim and certifying the repair. The correction gets no ritual, no dateline, no photograph, and above all no institutional owner whose actual job is to make anyone hear it.

A journal retraction reaches a fraction of the readership of the paper it retracts. A patched vulnerability never finds the people who read the disclosure. A recall runs at the bottom of page nine. The correction runs on A17.

So the public ledger is one-directional by construction. Every failure enters the record with full production values. Every fix enters as a PDF.

Trust does not decay because publics are stubborn or innumerate. It decays because there is no institution whose job is to announce that something got fixed — and so the only entries anyone ever sees are the debits.

v · three years later

All of the above was written on the day the valve opened. The thesis made a prediction, which means it can be checked, and three years is now enough data to check it.

TEPCO has discharged roughly 163,800 cubic metres — a bit over a tenth of the total — across more than twenty batches. Every one has been sampled, verified by the IAEA's task force, and published. Tritium has come in far below the 1,500-becquerel operational limit each time. Three years of clean values, exactly as predicted, entering the public record exactly as predicted: as a PDF.

The integral did move, a little. China's ban held for just over two years. The first Japanese scallops — 6.6 tonnes, out of Hokkaido — cleared Chinese customs in November 2025, with Fukushima and nine neighbouring prefectures still excluded. Slow, partial, conditional, real.

Two weeks later, Beijing suspended imports again. The trigger was the new prime minister's remark in the Diet that a Chinese attack on Taiwan could warrant a Japanese military response. The stated reason was the need to monitor Fukushima's treated water. A foreign ministry spokesperson added that there was, in any case, "no longer a market for Japanese seafood."

Which resolves the argument the 2023 coverage could not resolve. Three years of verified data did not move the ban, because the ban was never a reading. It was a lever. The water was the handle, and the handle stayed bolted on, available for use in a dispute about Taiwan that had not been started yet.

The engineering question got answered. It was always the answerable one — chemistry responds when you ask it properly, with instruments, in public, at length. Japan answered it thoroughly and slightly too late, and then spent a decade discovering that a correct answer to the wrong question is worth approximately nothing at the fish market.

We know exactly how to build a purification system for water. We spent twelve years and a fortune on this one, it works, and it will keep working until 2053. We have built nothing at all for the second substance — the one every institution on earth generates in far greater volume, the one with no filtration stage, no discharge schedule, no independent task force, and no dashboard.

Those tanks are also full. Nobody has proposed draining them.