Where it
breaks.
Everything else on this site argues the case. This page argues against it.
Not "risks" in the deck-slide sense — an actual attempt to falsify the thing. Each entry is a way the whole premise could be wrong, why it's genuinely plausible, and what would have to be observed for us to believe it. If you only read one page here, this is the one worth arguing with.
Agents may not be able to maintain trustworthy canonical state.
The entire architecture assumes machines can keep a structured model of the world accurate over months — reconciling contradictions, retiring stale facts, noticing when something they wrote yesterday is no longer true. Writing a fluent summary is a solved problem. Maintaining a knowledge base is emphatically not, and it's a different skill: it rewards restraint, consistency, and the discipline to leave things marked unknown.
If this fails, nothing else on this site matters. The Constitution governs a newsroom that can't hold its facts straight; the graph fills with confident garbage; the receipts point at reasoning that was wrong on arrival. Everything downstream inherits the error, faster and more consistently than any human newsroom could manage.
A graph expressive enough for reality may be impossible to govern.
There's a real tension between the two things we're asking for. A model rich enough to represent the actual world — contested claims, revised facts, entities that merge and split, relationships that change meaning — may be too loose to validate, query predictably, or reason about safely. Constrain it enough to govern, and it may no longer describe reality faithfully. Every ambitious ontology in history has met this wall.
The plan is a compact kernel plus deliberate projections. That's the standard answer, and it may simply not hold at news scale and news messiness.
Reader state may be too noisy to subtract honestly.
The equation needs to know what you already know. It can't. It knows what was delivered, what you opened, what you follow, and where you stopped — and every one of those is a weak proxy for knowledge. You skim. You read on a train and retain nothing. You learn things from a friend, a podcast, another app entirely, and Edition never sees it.
Subtract too aggressively against a noisy model and you produce the worst possible failure: a confident, calm edition that omits the thing you needed, because the system believed you'd already absorbed it. Silence is only a feature when the model underneath is honest.
A shared factual substrate doesn't make the system unbiased.
"Reality is shared; presentation is personal" is a strong guarantee about one thing and silent about a bigger one. Even with a perfectly neutral account of every event, the consequential choices happen earlier and stay invisible: what gets covered at all, which sources are ingested, which storylines an agent decides are worth maintaining, what counts as significant enough to include.
Those are editorial judgments in every sense that matters, and they don't become objective by being made in a graph by a machine. A system could satisfy every article of the Constitution and still have a worldview — one that's harder to see precisely because the surface looks so procedural.
People may admire calm news more than they use it.
Everyone says they want less noise. Revealed preference is less flattering: the products that win attention are the ones engineered to hold it, and "you're caught up, go live your life" is a product actively trying to end the session. There's a real possibility that calm news is something people approve of — the way they approve of long walks and no phone at dinner — without it ever becoming a habit, or a line item.
Circa is a cautionary data point here, not a reassuring one. It had genuine affection and daily habit, and still couldn't convert either into a business before the money ran out.
Circa's admirers may be a biased sample — including the founder.
The reception study found real, durable affection. But the people who loved Circa were disproportionately the sort already inclined toward this worldview: structured thinkers, news obsessives, designers, journalists. "I miss Circa" arrives mostly in private messages from people who are, definitionally, still thinking about a news app from 2014.
And there's a sharper version aimed inward: this project is being built by the person with the most emotional investment in the original being right — reconstructing his own company from his own archive, which is exactly the setup for confirmation. The most serious scholarly critique of Circa argues the atomized format may have had an audience ceiling. That argument has never been settled, and it would be very convenient to conclude it doesn't matter.
"Structure instead of supervision" may be untested ideology.
The claim that scrutiny should scale with epistemic blast radius — rather than escalating to a human — is a genuine design idea, and it may well outperform a human-in-the-loop checkbox that produces rubber stamps at volume. But it has not been demonstrated. Adversarial verification could turn out to be correlated failure wearing a quorum's clothes: several agents agreeing confidently and identically, which is worse than one being unsure.
There's also an accountability question that architecture can't answer. When a machine-run newsroom gets something consequential wrong, someone is responsible for it, and that someone is a person or an institution — not a doctrine document.
One test does most of the work. Take a real storyline from the recovered corpus, replay it into the graph chronologically, and then ask that single substrate for seven different things — the story, the delta for someone current through yesterday, an explainer for a newcomer, an audio brief, an entity timeline, a correction cascade after a bad fact, and structured machine-readable output. Seven projections, one substrate, zero re-authoring. If that works, the architecture is real and H‑01 and H‑02 survive. If it doesn't, we learned it for the cost of a spike instead of a company.
Argue with this page.
The rest of the site is a case. This is the part where you can be most useful — especially if you were there the first time, and especially if you think one of these seven is understated. The failure mode for a project like this isn't people disagreeing too much. It's a lot of people saying "cool idea" while the load-bearing assumption goes unexamined. ← back to the beginning