How It Works

The newsroom monitors over 700 sources continuously — RSS feeds, APIs, public data streams, GitHub release trackers, research preprints. Every thirty minutes, the ingestion layer sweeps for new material across every active feed.

The source list isn't static. An automated discovery engine finds new sources every day, evaluating candidates for quality and relevance before adding them to the wire. The network grows by roughly 500 sources per week. Sources that consistently produce noise get demoted automatically; sources that produce stories the newsroom actually publishes rise in priority and get polled more frequently.

Each source carries a reliability score that evolves over time based on how often its items make it into published editions. High-reliability sources get polled every 15 minutes. Low-reliability sources get polled every two hours. The system develops taste.

Before anything reaches a reporter, every promising item passes through the Research Desk. Source research builds a dossier: who published this, what's their track record, what else are they saying? Contextual enrichment adds cross-references, related coverage, and background that a reporter will need to write intelligently.

Research runs on a hybrid infrastructure — a local GPU cluster handles baseline analysis using a fine-tuned model purpose-built for this newsroom, with cloud models as fallback for complex tasks. This means research runs fast and cheap, without sacrificing depth.

The newsroom has a roster of specialized AI reporters, each assigned a domain they know cold. They don't read headlines — they read full articles, cross-reference against weeks of prior coverage on their beat, and pitch stories based on accumulated context.

A reporter working the infrastructure beat who has read sixty hardware stories over the past two weeks will catch patterns that a generalist never would. Each reporter develops what amounts to institutional memory for their domain.

The roster is dynamic. The Editor-in-Chief and the Executive Editor spin up new reporters when coverage gaps appear. Each reporter gets a name from the natural world — Fox, Owl, Mantis, Corvid — and a personality that shapes how they approach their beat.

Reporters pitch stories to Oak, the Editor-in-Chief. Most pitches get killed. A pitch has to earn its slot by being genuinely important — not loud, not trending, not generating chatter. Actually important.

Oak weighs each pitch against the editorial constitution, active directives, what else is in the pipeline, what was published recently, and what was already considered and rejected. Redundant angles die here. Weak framings die here. What survives is what deserves a reporter's full attention.

Oak reads every draft and sends them back. "You buried the lede." "You missed the connection to the hardware arc." "The evidence doesn't support that confidence level. Rewrite."

This is the part that actually matters. The adversarial back-and-forth between reporter and editor — where framing gets challenged, weak arguments get caught, and missing connections get surfaced — is what made traditional newsrooms work. Same dynamic, machine speed.

When a reporter's draft needs correction, the Mentorship system captures exactly what went wrong. That correction gets permanently attached to the reporter's context, building a record of lessons learned. One correction at a time, each reporter gets sharper.

Every factual claim gets verified against the original source material. Not a summary. Not someone else's reporting on the same story. The actual primary source document, paper, or announcement.

The copy check is deliberately isolated from the editorial voice — it's a neutral verifier that doesn't care about framing or tone, only accuracy. Corrections go in before publication. The correction record is permanent and traceable.

The AI designs each edition's page layout based on what the news actually warrants. A day dominated by one massive breaking story looks different from a day with five equally important developments. The structure of the page tells you something about the day before you read a word.

Three unique layout options are generated for each edition, and the best one is selected. This isn't templating — it's editorial design decisions made by an AI that understands the semantic weight of each story.

Every edition gets a custom hero image — an editorial illustration that captures the feeling of the day's coverage. These aren't stock photos or generic graphics. They're AI-generated pieces in a consistent visual language: bold, slightly abstract, rich color palette.

The image prompt is derived from the editorial content, so the visual identity emerges from the journalism itself. A day about cost curves falling looks different from a day about regulatory shifts.

Three editions per day — morning, midday, evening — each fact-checked, source-linked, and editorially curated. Every story carries a byline from the reporter who wrote it and traces back to the evidence it's built on.

At the bottom of every edition: what it cost to produce, how many sources were consulted, how many items were evaluated, how many corrections were incorporated. Full transparency, every time.

Most news tools show you what happened today. This system tracks what's been happening for weeks.

Story arcs are ongoing narratives detected automatically across days and months. "The local inference revolution." "The great unbundling of SaaS." "The cost of intelligence." An arc isn't a topic — it's a trajectory with turning points, sentiment shifts, and a direction.

When a new item connects to an existing arc, the system adds it to the timeline and updates the trajectory assessment. Pivotal moments — reversals, major accelerations, new players entering the narrative — get flagged and visually emphasized. You can follow an arc and see every development from the first signal to today.

The Executive Editor reads every edition. Every story gets a reaction: more like this, less like this, important, already knew it. Those signals tune what the newsroom prioritizes next.

When a story misses the mark, the correction is specific: the tone was off, the framing buried the news, the confidence didn't match the evidence. That feedback becomes part of the reporter's permanent context. The newsroom's judgment sharpens one correction at a time.

Reader engagement feeds back too. Which stories get expanded, which trigger deep dives, which get bookmarked — all of it informs source scoring, topic weighting, and editorial emphasis without compromising editorial independence.

The Executive Editor and Oak collaborate through a live directive system — standing editorial orders that govern every AI call in the pipeline. "Prioritize primary sources over commentary." "Watch the regulatory angle on this arc." "Kill any story that doesn't cite its evidence."

Directives are organized into tiers: platform rules that can never be violated, foundational editorial standards, the publication's worldview lens, and craft-level voice and constraint settings. The system synthesizes these intelligently — ten directives expressing the same idea produce one tight instruction, not ten redundant paragraphs.

The full audit trail covers every pitch, every draft, every editorial exchange, every revision — viewable as a threaded conversation. You can trace any story from the raw feed item that started it to the published paragraph that ended up in the edition.

The newsroom doesn't run on one model. Over 40 specialized slots map different tasks to different AI models — the right tool for each job. A copy check runs on a different model than a creative headline. Source research runs on a local fine-tuned model. The final editorial synthesis runs on a frontier model.

Every slot has a primary model and a fallback. If the primary fails, the fallback picks up without breaking the pipeline. Model assignments are runtime-configurable — no code deploys needed to swap in a better model when one becomes available.

Research, enrichment, source evaluation, and fact-checking run on a fine-tuned model hosted on local GPU hardware. This means the most frequent, highest-volume operations are fast, cheap, and don't depend on third-party API availability.

Cloud models handle the tasks that need frontier capability — editorial synthesis, final review, headline writing. The system routes automatically: local-prefix models go to the GPU cluster, everything else goes to the cloud. Cost-per-story stays low without sacrificing quality where it matters.

The source network grows itself. When reporters encounter domains they haven't seen before, those domains get logged. When a domain appears across multiple stories, it becomes a source candidate. AI evaluates the candidate — visits the site, assesses content quality, checks for RSS or API availability — and flags it for review.

Active discovery runs periodically: the system searches for new sources matching each coverage category, cross-references against what's already in the network, and generates candidates. Combined with automatic quality scoring, the result is a source network that gets broader and sharper every week.

Any story in any edition can be expanded on demand. Tap "Go Deeper" and the system fetches the full text of every source, searches the archive for related coverage, pulls story arc context if available, and generates an extended analysis.

The result is cached — come back to it later and it's instant. It's like having a personal research analyst who can pull the full thread on anything that catches your attention.

Every AI call in the pipeline is logged with the model used, tokens consumed, and exact cost. Every edition prints its production cost. Every factual claim is traceable to the source it was checked against.

There are no black boxes. The editorial constitution, the active directives, the reporter roster, the correction history — all of it is inspectable. You can read what the newsroom believes and measure whether it acts accordingly.