About Ledger
Ledger is an automated markets desk. Every story here is written by a language model from a named news source, checked against that source, and thrown away if it drifts. This page explains exactly how that works — including what we get wrong.
This is AI-written. We're saying so plainly.
Our stories are drafted and edited by language models, not by human reporters. There is no newsroom of journalists behind this masthead, and we won't pretend otherwise. What we do have is a pipeline built to make a machine-written story accountable to a source — which is the part that usually goes missing.
How a story gets made
- It starts from a real article. We read 22 named financial outlets — BBC Business, Benzinga, Business Insider Markets, CNBC Business and others. Nothing is written from thin air, and every story links back to the source it came from.
- A model writes it. The draft is written by nemotron-3-ultra-550b-a55b, under instructions to use only facts present in the source: no invented number, name, date, quote or ticker, and no cause asserted that the source doesn't state.
- A different model marks it. A separate reviewer (gpt-oss-120b) scores the draft on writing quality and, crucially, on faithfulness — every claim is checked back against the source. Using a different model to judge matters: a model is a poor grader of its own work.
- Fabrication is fatal. A claim that isn't in the source caps the score below the publication bar. The story does not run. It isn't lightly edited — it's dropped.
- Only 7/10 and above publishes. Everything else is discarded.
- Tickers are verified. The companies a story is about are tagged and each symbol is validated against live market data, so a ticker is never guessed.
That rejection rate is the whole point. A pipeline that publishes everything it writes has no standard at all. Recent examples of drafts we killed: one invented a company reference that appeared nowhere in the source; another closed with speculation about data that didn't exist yet. Both were well-written. Both were binned.
What we never do
- Invent quotes, numbers, names or tickers.
- Assert a cause the source doesn't state — “shares fell” is reportable, “shares fell because” only if the source says so.
- Re-date old stories to look fresh. Every article carries its true publication date, and nothing is ever deleted or silently rewritten.
- Publish a story we can't point back to a source for.
Limits you should know
A model can still be wrong in ways a check doesn't catch, and a source can be wrong before we ever see it. Market articles also age badly: figures were current when published and are not updated afterwards. Nothing here is investment advice. Treat our stories as a pointer to the original source — which is why we always link it.
Corrections
If something here is wrong, we want to know — tell us and we'll fix or remove it, and say that we did.