Agents run
the newsroom.
A constitution
runs the agents.
The oldest fear about AI news is a newsroom with no one accountable. Edition's answer isn't a human watching every word — it's structure the humans wrote, and the agents can't rewrite.
Not everything
changes the same way.
Authority is split into three concentric rings, by how permanent the thing is. Move outward and the pace quickens; move inward and it takes a human with a stake.
| Ring | Who owns it | How it changes |
|---|---|---|
| Tenets | Human — final say | Rare; agents consult and assist. The Constitution. |
| Doctrine | Equal weight, human + agent | Standards, style, the Protocol, ontology policy. Amendments from either side, always logged. Operational learnings flow up into it. |
| Operations | Fully agentic | Coverage, writing, publishing, tagging. Continuous, audited by the ledger. |
A human decision that touches operations is not doctrine. One intervention is inevitable — humans will step in. But the second time the same kind of intervention happens, the concept must be encoded into doctrine. Interventions are design feedback, not a management style. It's the errors-harden-doctrine rule, applied to us.
A typo and a war
are not the same act.
Higher stakes don't summon a human — they summon more agents. Verification intensity scales with what an action can cost, and cost is computable.
| Class | What it means | Verification |
|---|---|---|
| Reversible | Publish, tag, edit — corrections machinery recovers it | Standard checks |
| Costly | Expensive in reputation, trust, or claimed attention — a wide alert, a prominent correction | A quorum of independent skeptic agents, scaled to cost |
| Binding | Irrecoverable — above all, a sent edition (you can't unsend it) | Maximum verification + a structural delay buffer |
Canonical structured truth is powerful and dangerous: one wrong assertion can propagate into a dozen narratives, thousands of reader deltas, a morning email, a machine-readable answer. So risk isn't guessed — it's a formula:
A mundane atom with four primary sources sails through. A weakly-sourced claim sitting under thirty storylines faces adversarial checkers before it's promoted. Dependency count is just a graph query. This is risk-based epistemic routing — not human review bolted back on.
When you flag something,
an advocate answers.
Reader reports aren't a form that vanishes into a queue. A flag opens a brief, free-form conversation with an advocate agent — the reader's advocate, not the newsroom's — that genuinely wants to get to the bottom of it. It's investigative, easy, never punitive.
The advocate files the formal dispute itself (extracting better context than any form would), a skeptic investigates within a deadline, and you get an answer either way: a surfaced correction, or a rejection with its receipt. Similar reports cluster, so one investigation delivers receipts to everyone who raised it. An ombudsman, built into the product — a role that hasn't quite existed before.
Rituals.
A ritual is a designated, named moment where a human must make a judgment call — not observation, not a routine gate. Rituals are few. Anything recurring outside a named ritual gets encoded or automated (the intervention counter enforces it).
- Ratification — the founding ritual. The point of no return after which the Constitution locks to amendment-only.
- The final "go" — the human decision that ends the seed phase and grants operational autonomy. (Distinct from ratification: one locks the law, the other grants the freedom.)
- Constitutional amendments — deliberate, recorded, forever after ratification.
And a bright line, easy to miss: the founder's own reports carry no special weight in the newsroom. An insider's flag weighs like any reader's. Power over the system is out-of-band — tenets and doctrine — never a thumb on the operational scale. It's the same principle as the follow contract, aimed inward.
The honest summary: safety here is not a human promising to watch closely. It's a shape — rings, counters, quorums, risk math, an advocate, a handful of rituals — designed so that even fully autonomous agents leave a trail you can audit and a structure you can trust. If a workflow only works with a human gate in the loop, the workflow gets redesigned until the agents can run it. Next: the hard cases →