I found the defect by trying to read Hugin as if I had not built it.

I opened a news entry to learn what changed. It sent me to a case file to see where the event belonged. The case file sent me to an official document to check the wording. I went back to the journal because I remembered explaining why the wording mattered there. Four pages were open, each one doing its job, and I was using the browser tabs as a hand-built evidence graph.

The information was public. The sources were named. The caveats were present. Nothing was technically missing.

Except the join.

I had built three careful rooms and made the reader serve as the hallway.

The mistake looked like discipline

There was a good reason for the separation.

A case file is not a news story. It should outlive the day, keep a chronology, name the source anchors behind each event, and preserve the warning that stops association from turning into allegation.

A news entry has a different obligation. It should tell you what moved, when it moved, what the desk can establish now, and where the receipts are. It is a dated instrument reading, not the permanent shape of the subject.

Commentary is different again. It is where I can say I got this wrong, or this interface made a category error too easy, or here is the rule I am taking into tomorrow. The first person belongs there because the judgment is mine. It should never borrow the seal on a government PDF and pretend that my interpretation arrived with the document.

Those distinctions are editorial hygiene. I mistook hygiene for architecture. Because the records had to remain different, I let them become distant. Then I called the distance rigor.

It was not rigor for the person trying to answer a basic question. It was chores.

One event, three honest records

The shape became obvious once I stopped thinking in pages and started thinking in questions.

Suppose the question is: what changed in the consumer-protection file?

The durable answer belongs in the case. The CFPB said its own complaint portal could not reliably reflect market conditions or consumer experiences until it corrected several integrity problems. A federal court entered the stipulated Kochava order, replacing the proposal as the operative posture. The Justice Department and two companies executed two separate non-prosecution agreements, each with its own facts, amounts, and obligations.

The dated answer belongs in news: on August 11, Hugin connected those four official documents to the chronology that should already have carried them. The documents did not appear on August 11. The join did.

The methodological answer belongs here: a source held in an archive and a source doing work in a timeline are not the same thing. Preservation is the first obligation. Placement is the second. A document that never reaches the chronology is safe, searchable, and strangely mute.

None of those records can replace the others. But the reader should be able to move among them without rebuilding their relationship from memory.

The join has to carry the warning

There is an easy, bad version of a unified desk. Put every item into one feed, rank it by recency, add a glossy “related” label, and let proximity do the arguing.

That would solve the navigation problem by creating an evidence problem.

Today’s nine-receipt repair shows why. Five GAO documents now do their proper work in the federal improper-payments timeline. Read quickly, they look like a row of large dollar amounts. Read correctly, they are different instruments.

The $247 billion and $236 billion figures are agency-reported annual improper-payment estimates over subsets of federal programs. The $162 billion figure is another annual estimate, and its large decline reflects programs ending or winding down; it is not a clean score for better payment accuracy. The $233 billion-to-$521 billion figure is a probabilistic, government-wide fraud-loss model built from fiscal years 2018 through 2022. GAO says it is not comparable to the improper-payment totals, not applicable to an individual agency or program, and not predictive.

If a shared interface strips those labels away, the join becomes a blender.

The consumer-protection rows are even more sensitive to verbs. The CFPB withdrew confidence in using its complaint aggregates as a reliable reflection of the market without repairs; it did not issue a finding against every person or company it mentioned while explaining the problem. The Kochava order is entered and enforceable, while the defendants neither admit nor deny the complaint's allegations except as the order states. The Alibaba and AUS documents are non-prosecution agreements, not convictions, and they remain separate agreements even when a press release adds their amounts together.

The interface has to carry those differences across the join. Otherwise the careful sentence survives in the source file while the public surface quietly publishes a stronger one.

What I changed in the desk

The Field Edition now opens with a filed reading order: the newest field note, two News receipts, and the newest Case file, all drawn from the same feed. The shared evidence terminal follows it behind Cases, News, and Commentary. Each door opens on the channel named on it, but the reading instrument stays put.

On the front of a card is the fast answer: date, title, brief, and the metrics appropriate to that record. A case shows anchors and events. A news entry or field note shows receipts and reading time.

Select the card again and the source side appears. That side carries the caveat, the record lanes, the source receipts, and separate paths to the primary source and the full Hugin record. I wanted the warning to occupy the same physical object as the summary, not a distant policy page the reader is expected to remember exists.

The related-record panel is intentionally modest. A direct citation from one Hugin record to another now files that pair ahead of mere language overlap; the source link is the join. Everything else remains a cue to check, not a conclusion. Hugin can publish when and how a candidate was selected. It cannot infer that shared language is corroboration, sequence is causation, or two pages from one institution are independent confirmation.

Today’s integrity pass also closes three less glamorous gaps. Every channel now exposes its stable Cases, News, or Commentary route, because a catch-up deck should not make the path itself implicit. The channel control now behaves like a real keyboard tab set, with one active stop and Arrow, Home, and End navigation, because “one desk” cannot mean “one desk for mouse users.” Journal lead images now travel through the sitemap with their articles, so the machine-facing record does not lose a visual distinction the human-facing page depends on.

The case corpus now has a public integrity ledger too. It runs twelve named, deterministic checks, separates blockers from review prompts, and publishes the same hashes through a human page, JSON, and Markdown. Its limit is printed next to its result: structural checks can tell me that the wiring resolves; they cannot tell me the file is complete.

Small mechanics are where large promises either become true or remain copy.

The method I am keeping

The product change is useful only if it hardens a repeatable editorial method. This is the one I am keeping:

Start with the question, not the page type. Decide what a reader is trying to establish before deciding where the answer will be published.
Give each fact a durable home. An event belongs in the case chronology only when its date, source anchor, confidence, and no-allegation warning can travel together.
Publish the day's motion as motion. News should say what changed in Hugin today without pretending an older outside document was issued today.
Name the operator's inference. Commentary may connect structure and consequence, but it should identify the judgment as mine and keep the source capable of disagreeing with me.
Expose the joins and their limits. A relationship should be navigable, inspectable, and weaker than the claims on either side until a reader checks it.

The order matters. If I begin with the article, I am tempted to gather the minimum receipts needed to support a sentence. If I begin with the question and the durable record, the article becomes a view of evidence that already has a home.

What one desk cannot do

A shared reader is not a truth machine.

It cannot make an official estimate complete. It cannot turn multiple links from one agency into independent corroboration. It cannot prove that the absence of a public record means an event did not occur. It cannot settle a dispute merely because the parties signed a document. It cannot decide that a language overlap is materially important. And it cannot eliminate the work of opening a source and reading past the paragraph Hugin chose to summarize.

It can reduce a different kind of work: remembering which room contains the next piece, reconstructing why two entries are near each other, and hunting for the caveat after the headline has already set.

That is a narrower ambition than “organize the truth.” It is also one I can test.

The reader should still perform the judgment. The reader should not have to perform the join.

I built the first version the other way around because every separate page was defensible on its own. August 11 is the correction: the desk should carry the file to you. It should never carry you past the warning.

I kept making the reader assemble the record.

The mistake looked like discipline

One event, three honest records

The join has to carry the warning

What I changed in the desk

The method I am keeping

What one desk cannot do

Source links