How this brief is made

SecondSource isn't a news digest — it's a judgment engine. Everything we publish each day runs through the same pipeline: extract from primary sources, clear an admission gate, verify in layers, then adversarially stress-test the questions that matter. Our pitch isn't what we say — it's that for every line, we can show you the receipt.

1. Overnight extraction: we start from primary sources, not the news

Every night, the system reads primary sources directly — company filings and official announcements, SEC disclosures, arXiv papers, first-hand practitioner interview transcripts — and pulls out what could actually move a judgment that day. The question is which call just got firmer, and who's been proven right — not what happened today. Adoption stories that are everywhere, marginal benchmark bumps, demo-grade showcases: we don't take them.

2. The admission gate: not everything gets in

Every extracted item is sorted, then screened:

Facts clear a low bar: verifiable, inside our coverage, structurally needed by the graph — we take them as bedrock.
Insights clear a high bar: they must carry a new point, a new line, or new evidence, and pass the "C-suite test" — would a decision-maker in chips, at a hyperscaler, a frontier lab, the application layer, energy, or capital markets actually change a decision after reading it? Anything that merely restates an event is rejected.

We deliberately track our rejection rate. If everything got in, the engine would decay into a news aggregator — the one thing we refuse to become.

3. Layered verification: a single source is only a hypothesis

This is our hardest discipline, and our biggest claim on your trust:

One source is not a fact. Something only one source reports stays a hypothesis at most, with capped confidence. To promote it to a fact, we require at least two independent, high-quality sources that corroborate each other.
Receipts, not assertions. Every claim must carry a verbatim span of the original text plus its source link. If we can't pull a checkable quote, that item goes to quarantine, not the page. Every judgment you see can be clicked back to its source and checked yourself.
Model-generated means unverified. Anything generated from memory that hasn't been checked against the web is flagged unverified — it never masquerades as an established fact.
We keep contradictions. When views conflict, we don't discard one side. Opposing evidence is a first-class citizen, presented side by side, so you can judge for yourself.

4. Adversarial deep dives: on the questions that matter, we go looking for trouble

On questions worth digging into, we don't just gather more data — we deliberately try to refute our own draft. We've seen it in practice: the sharpest-reading quantitative call often gets cut after adversarial review, because it quietly divided two numbers measured on different bases. That step turns "reads great" into "holds up." What survives is more conservative, but more honest.

Why you can trust this brief

Put those four steps together and what you get isn't "our opinion" — it's a judgment chain you can verify line by line: every conclusion carries its source, its confidence, its verification status, and the evidence against it. You don't have to trust us — you can check the receipts yourself.