PAN

AN started as a research question at Parallax AI: what happens if you force three AI agents to deliberately disagree about the same news event — and then make a fourth one reconcile them in writing? How it works: Every story goes through five passes before publication: Positive read — an agent steelmans the most generous interpretation Neutral read — strictly factual, dateline-style, evaluative language forbidden Negative read — a good-faith red-team pass that surfaces what the press release doesn't want you to ask Cross-read — a fourth agent resolves factual conflicts between the three, against source Synthesis — a fifth agent merges everything into one article The three reads stay visible in the sidebar, with the opinion balance shown as percentages (e.g. +30% / =45% / −25%). No human editor touches the output. Wall time from ingest to publish: typically 90–180 seconds. What surprised us after 18 months of internal beta: the output is more cautious than a single-author newsroom — it admits uncertainty by construction and surfaces conflicts a single writer would tend to flatten. We honestly expected it to be more boring. It isn't. What PAN is not: Not a wire service — we don't break news, we synthesize it. Our value is the read at 06:14, not 06:01. Not an opinion outlet — there's no editorial column. The synthesis is constrained to reconcile, not argue. Not anonymous — every article carries its pipeline version number, full source citations, and a note that no human reviewed it. The missing byline is a claim, not an evasion. PAN is one of four public services built on the same multi-perspective method — the others cover historical events, legal decisions, and claim-by-claim fact verification. I'd love your honest takes on: Do the three visible reads actually change how you trust (or distrust) an article? Is "no human editor" a feature or a bug to you? Genuinely curious where the line is. Corrections are automated within 6 hours — and yes, that applies to feedback from this thread too. 🙂

Discussion (0)