ellipsys
·9 min read·Media Systems

Full Disclosure

Filed byThe Public Record Office

On January 30, 2026, the Justice Department published over three million additional pages responsive to the Epstein Files Transparency Act. Combined with earlier releases, the Department put the total production at nearly 3.5 million pages, along with more than 2,000 videos and 180,000 images. Material not produced, the Department said, was duplicative, privileged, excepted under the Act, or unrelated to the case.

One reader, spending one minute per page, working a forty-hour week and taking two weeks off, finishes in about twenty-nine years. That is arithmetic, not research. It is offered because 3.5 million is not a quantity anyone can hold, and twenty-nine years is.

The Department reports that more than 500 attorneys and reviewers worked on the production. A member of the public is issued one reader.

We record this not as a scandal but as a filing. The release was compliance. It satisfied a statute. Every page of it is public, permanently, at no cost, to anyone — and unreadable, in the only sense that matters to a reader.

Published and unreadable used to be a contradiction. It is now a technique. It has more than one form, and the forms are worth separating, because they are usually discussed as one thing and only one of them requires intent.

Three ways volume conceals

1. Rate. The original and best-documented form. Steve Bannon, to PBS Frontline: "The opposition party is the media. And the media can only — because they're dumb and they're lazy — they can only focus on one thing at a time." The remedy follows immediately. "All we have to do is flood the zone. Every day we hit them with three things. They'll bite on one, and we'll get all of our stuff done, bang, bang, bang. These guys will never — will never be able to recover. But we've got to start with muzzle velocity."

The phrase usually quoted is flood the zone. The operative phrase is muzzle velocity: not the content of any round, but the rate of fire. Nothing is hidden. The cycle cannot metabolize items as fast as they arrive, and an item that is not metabolized is functionally an item that did not occur.

2. Mass. Everything at once, undifferentiated, in a shape that resists reading — the scan of a fax, the image-only PDF, the search that returns all of it, sorted by date. The Epstein production is the specimen. Note the property that makes it durable: it is not a failure of transparency and cannot be attacked as one. Full compliance and full concealment have converged on the same physical act, and the party performing it never has to say which one it was doing.

3. Dilution. The signal is present, correctly filed, at a concentration too low to find. The Justice Department's summary of fiscal 2024 agency reports under the Freedom of Information Act records a government-wide high of 1,501,432 requests received, a 25% rise in a single year, and a backlog up a third, to 267,056 pending. Every figure there describes a system working harder. None of them describe a reader who found the transparency they were looking for.

Only the first requires a strategy. The third requires nothing at all. All three rest on a single assumption, and it is the only one they make: that reading is done by people, at people speed, out of a fixed daily budget. Each is an attack on that budget. None is an attack on the record.

The budget moved

Since 2019, a coalition of more than forty California newsrooms has filed public-records requests with nearly 700 agencies — police, sheriffs, prisons, coroners, oversight boards — under the state's police transparency laws. The returns arrived as returns always arrive: PDFs, audio, video, scattered across hard drives and newsroom laptops. KQED, the project's coordinating hub, reports that the archive now holds more than 22 terabytes, and that roughly 1.5 million pages are public in a searchable database.

Machine reading does the work no funder would pay a human to do: cluster documents into cases, extract incident dates and officer names, index everything. Then the part that matters. The project cross-references its own holdings against the use-of-force and in-custody-death data that agencies are separately required to report to the state. Where the state data shows an incident and the records do not, the gap is flagged and pursued. By this method, KQED reports, the team has identified thousands of additional cases that agencies might not otherwise have sent.

Read that mechanism again, because it inverts the thing. The concealment was an absence — records legally required and quietly not produced. An absence is invisible against a partial corpus and visible against a complete one. The mass that made the archive unreadable is what made the hole in it detectable. Volume stopped being the adversary's asset the moment something could read all of it.

The configuration

This publication has spent its existence describing summarization as subtraction — the thing that arrives smaller and files the difference as efficiency. We are not revising that. We are recording the one case where the arithmetic runs the other way.

Subtraction needs a quantity to work on. Compression removes signal when signal was getting through, and against 3.5 million pages nothing is getting through. The baseline is zero. A summary of an unreadable record cannot take away what no one could reach, so whatever structure it adds is the first structure there has been. That is the whole difference between summarizing a memo and summarizing a flood, and it is a difference in the record, not in the technique.

Which is exactly why the technique has to be right, and the measurements are not comfortable. Dahl, Magesh, Suzgun and Ho, in the Journal of Legal Analysis, put hallucination rates between 58% for GPT-4 and 88% for Llama 2 on specific, verifiable questions about randomly selected federal court cases; the models also struggle to predict their own hallucinations and tend to accept a user's incorrect legal premise rather than correct it. An evaluation of locally-run models for investigative document search, presented at the 2025 Computation + Journalism Symposium, found high citation validity in systems that still propagated errors through multi-stage synthesis — a mistake made while summarizing arrives at the conclusion wearing a citation.

None of that is an argument against doing it. It is a specification, and the California project reads like one. Machine reading sorts, links, extracts, and summarizes. Search returns literal matches from the text of the original government record. Every case the system flags as misconduct is reviewed by a person. Model output is checked against hand-built samples drawn from every year and every category of agency in the collection. The site tells users plainly to treat search as a starting point and to read the underlying documents before reaching conclusions. Those are procedures. They were designed, they cost money, and they are why the summarizing is load-bearing rather than decorative.

So the finding is not that indexes are the answer and summaries are the problem. Alone, neither is either.

A summary on its own is an assertion. The rates above are why that matters: something wrong on a checkable question between 58% and 88% of the time does not get taken at its word. Indexing, extraction, sorting and search have the opposite defect. They are decades old, they are exact, and they have never once made a document dump readable — retrieval returns every literal match and no reason to open any of them. One technique is legible and cannot be trusted. The other is trustworthy and cannot be read.

Paired, each covers what the other cannot do. The summary makes the pile enterable. The citation makes the summary falsifiable. Every claim arrives carrying the location of the page that would contradict it, so nothing has to be believed and everything can be checked. A good summary of a 400-page use-of-force investigation, sitting on top of the 400 pages and one move from them, is the most valuable object in the archive — it is what lets one reporter cover 700 agencies instead of one. Cut the citation and the same summary is a smaller flood with better manners.

An index says it is here. A summary says this is what it says. Everything turns on whether the second sentence is still holding on to the first.

Filed

This is the one arrangement in which compression is aimed at an obstruction rather than at a signal. It works. It works for as long as somebody keeps paying to maintain the way back, which is a recurring cost with no constituency.

Nothing in the last two years suggests that bill keeps getting paid. The pressure runs the other way — toward the summary that arrives finished, toward the answer that saves the reading, toward 3.5 million pages compressed into a paragraph that nobody will check against 3.5 million pages.

The record was never the problem. The reading was.

Sources

A report about pointing back to the record is obliged to point back. Everything above is checkable here.

— Filed by The Public Record Office