How we actually reason.

Frontier AI, human judgment.

Before any prose gets written, we assemble a dossier: market data and on-chain history from Polymarket, news flow, the live conversation the price is reacting to, and the primary sources any of it points back to. Each piece is pulled through a dated, verifiable channel — not a free-form web crawl — so every item in the corpus has a timestamp and a provenance. The volume is larger than any single reader can hold in their head; that's exactly why we built the pipeline this way.

We then point frontier AI at that dossier. The models do the heavy lifting — reading the full corpus, linking claims to sources, flagging contradictions, and drafting a first-pass directional view. Human editors do the judgment call: deciding which thread actually matters, which argument holds up under pressure, and whether the evidence is strong enough for the desk to stake a side in print. Neither half works alone.

The model never searches the live web during writing. It reasons over a fixed, pre-fetched corpus. That makes the pipeline deterministic, cache-friendly, and auditable — and it means a rerun next week on the same corpus will produce the same shape of output.

Every claim traces to a source.

In the prose of every article you'll see inline tokens that look like [comment_1234] and [tweet_5678]. On the article page, those render as clickable source cards with the original text, handle, and timestamp. They are not decoration — they are the unit of evidence.

Between the research pass and the writing pass, a token validator scans the body and strips any citation ID that was not in the pre-fetched corpus. The model is physically unable to invent a citation. It can forget to cite something real, but it cannot manufacture a source that doesn't exist. If a claim is in the article, the receipt is two clicks away.

Opinions, not forecasts.

We do not publish Brier scores. We do not publish calibration curves. We do not publish point probabilities that we cannot defend in plain language. What we publish is a direction — a BUY_YES or BUY_NO call — and a short editorial that justifies it in human words.

This is deliberate. The honest answer to "what is the probability Iran's regime falls by June 30" is a shrug wrapped in a story, not 0.19. If you want the shrug without the story, Polymarket has the live price already. If you want the story, stay here.

Living articles. Transparent updates.

Every market gets one permalink for as long as it's open. When the evidence moves, the article is rewritten in place with a clear marker: the new "Current view" at the top, the previous view archived below it as "Previous view — archived," and a dated entry added to the history log at the bottom.

Nothing is silently rewritten. Every direction change the desk makes — with date, old signal, new signal, and the trigger — lands on the flip log. A view that moves when the evidence does is a feature, not a retraction: the changelog is there so readers can see exactly how it moved.

What we don't do.

We don't give financial advice — see the terms page. We don't aggregate other people's forecasts. We don't run paid sponsorships inside articles. We don't refuse to cover a market because the call went against people we like. We don't pretend the model is reasoning more carefully than a human reading the same five tweets and six comments would. And we don't publish numbers we can't reproduce from the exact same pre-fetched corpus on any machine.

Everything we do publish, we publish in one place, under one name, with the sources attached. That's the whole methodology.