The day started the same way for every reporter on the desk. Someone opened outlet after outlet by hand, mentally noting what was new against what they remembered airing days or weeks earlier. There was no single view of either. A story that looked fresh might already have run. A story that hadn't run might not be new at all, just a slow follow to something covered the night before.
Once a reporter had a candidate, the harder part started. They'd pitch it at the morning meeting, and executives would send it back if the angle didn't fit how the desk covers a story, forcing another pass through the same sources to find a better one. That back and forth happened most mornings, and it cost the executives time too. Neither side had a fast way to check a pitch against what the desk actually tends to run.
"I used to open ten tabs before I even knew if anything was new," Stella Oh, International Desk Reporter, said.
Stella was one of the reporters who walked through her actual morning while the problem was being audited.
The real problem wasn't finding foreign news. The problem was time: how long it took, every single morning, to work through that coverage by hand before anyone could sit in a room and decide what to run.
Overnight issue sweep. Each morning, the tool checks dozens of US and international outlets for the past 24 hours, groups articles covering the same event into a single issue, and checks each one against the desk's own broadcast history to flag it as new or a follow-up. Every issue comes with a brief, the source articles, and any related domestic coverage, all translated into Korean so nothing requires a second pass just to read it.
Angle recommendations. For each issue, the tool suggests an angle that matches how the desk actually covers stories, tuned on the issues the desk has historically chosen to put on air and on interviews with reporters like Stella about what makes a pitch land. It's not guessing at a house style from the outside. It's trained on the desk's own record.
Rough drafts on request. If a reporter picks a recommended angle, the tool writes a rough draft in the desk's style, pulling from the selected source articles. That wasn't part of the original ask. It came later, once the first two pieces were working.
The desk chose not to build real-time breaking alerts. An early version re-ran the full search every few minutes per open session, and the cost didn't justify it for a desk where genuine breaking news is rare. A slower, cheaper version stayed on the list for later.
The first version went live within a day of the initial conversation with Stella and her colleagues. Over the weeks that followed it grew from a plain list of headlines into the ranked, cross-referenced version that now runs on its own before the desk gets to the office.
The morning scan that used to take an hour is now instant. The tool runs before reporters arrive, so the review that used to open every morning is already done by the time they sit down.
Pitches get approved faster. Because each issue already comes with an angle scored against what the desk has actually run before, fewer pitches bounce back at the morning meeting, and executives spend less time sending reporters back to try again.
Reporters catch stories they would have otherwise missed. Scanning dozens of sources by hand meant defaulting to the same handful of outlets out of habit. The wider daily sweep has surfaced international stories no domestic outlet had picked up yet, stories the desk was then first to air.
Verifying a story's original source and building the final draft is still done by hand, and likely to stay that way. What's next is extending the same search to cover breaking alerts during the day, not just the overnight sweep, in a form cheap enough to leave running.
"It's learned what we've actually put on air, so the picks make sense before I even open them," Stella said.
Ocean's Halo ran its packaging supply chain across spreadsheets owned by three companies, retyped by hand at every handoff. One dashboard now shows what needs attention, and the plants and the supplier enter their own numbers.
Every operation has documentation that stopped being true at some point. This is a method for finding that point: extracting entities from records, building a graph of what should be, and reading the divergences against what is.
The pilot worked. A year later nothing is in production. The failure is almost never model quality. It is data nobody owns, permissions nobody granted, and a workflow nobody changed.
