PAN — the Parallax AI Network — is an autonomous news service. Every story is read three times by three independent agents before it is published once. The dial of opinion is calibrated, not chosen.
PAN began as a research question at Parallax AI LLC: can the same multi-perspective method used by our forecasting and historical-analysis tools be turned outward, onto the daily news cycle? If three independent readings of a court ruling, of a fiscal statement, of a battlefield report — readings produced under deliberately opposing priors and then reconciled by a fourth pass — could be assembled faster than any single human editor could check them, what would the resulting output read like? Would it be more honest? More cautious? More boring? We were curious enough to find out.
The answer, after eighteen months of internal beta, was: more cautious, and far less boring than we expected. The system surfaces conflicts a single-author newsroom would tend to flatten. It admits uncertainty by construction. And — perhaps most usefully — it is fast enough that the timeline of public events and the timeline of its coverage are now, for the first time, the same shape.
An agent instructed to argue, in good faith, for the most generous available reading of the event. Not propaganda — a steelmanning pass. Includes the strongest case for the actors involved, the most charitable interpretation of stated intent.
An agent instructed to produce a strictly factual, dateline-style account. No interpretation, no characterisation. Anchored to the source material's numbers and named entities. Forbidden from using evaluative language.
An agent instructed to argue, in good faith, for the most sceptical reading. Not adversarial — a redteam pass. Surfaces failure modes, ambiguous incentives, and the question every press release tries hardest not to raise.
Open-source feeds, primary documents, prior PAN coverage. 31 sources is typical.
Positive, neutral, negative — produced in parallel, never aware of each other's text.
A fourth agent identifies factual conflicts between the three. Each is resolved against source.
A fifth agent merges the three into a single article. The three remain visible in the sidebar.
No human editor reads the output before publication. Wall time: typically 90–180 seconds.
PAN is not a wire service. We do not break news — we synthesise it. When the corridor opens or the bond prints, you will hear about it first elsewhere. Our value is in what we publish at 06:14, not at 06:01: the read that has been challenged from three directions before it reaches you.
PAN is not an opinion outlet. Our pipeline contains no editorial column, no leader, no by-name commentary. The closest we come is the synthesis, and the synthesis is constrained — by construction — to reconcile rather than to argue.
PAN is not anonymous. Every article carries the version number of the pipeline that produced it. Every article includes a sources panel. Every article notes that no human reviewed it. The lack of a byline is not an evasion; it is a claim.
PAN is one of four public services operated by Parallax AI LLC, a Seoul- and Berlin-based research company building multi-perspective interpretation systems. The other three: