How this brief is made

This page covers two things: how we actually make the daily brief, and how you can check our work. The headline goes first — SecondSource is generated by AI, filtered through admission gates, and ruled on by a human at the points that matter. We don't ask you to trust it. We ask you to verify it.

One judgment's path, from intake to payout

Flow: 1. Primary sources -> 2. Admission gate -> 3. Two sources, one receipt -> 4. Adversarial review -> 5. Verdict card -> 6. Payout day. Each stage lists what gets blocked: stories everyone already has, mere restatements, unquotable claims, calls that don't survive rebuttal. — One judgment's full path, from intake to payout. The grey text beside each stage is what that stage blocks — what we actually publish each day is what survives all six gates.

Six gates, and every one of them throws something away

1. Primary sources: Every night we read primary sources directly — company filings and official announcements, SEC disclosures, arXiv papers, first-hand practitioner interview transcripts. The question is which call just got firmer and who's been proven right, not what happened today.
2. Admission gate: Facts clear a low bar (verifiable, in-coverage, structurally needed — they get in). Calls clear a high bar: would a decision-maker at a chipmaker, a hyperscaler, a frontier lab, the application layer, energy, or capital markets actually change a decision after reading this? If we can't answer that, it doesn't get in. We log every admit/reject call — 3,927 so far, and more get blocked or parked than get admitted.
3. Two sources, one receipt: Every claim has to carry a verbatim span of the original text plus a source link. Something only one source says stays in the fact layer, flagged single-source, confidence capped — it never gets written up as a call. That's why, across 435 live calls on this site, the highest confidence score is 0.8 — not one is written as a certainty.
4. Adversarial review: On questions worth digging into, we don't go get more data — we go try to refute our own draft. What survives is more conservative than the first cut, but it holds up.
5. Verdict card: Every forward-looking call becomes a card — who said it, the receipt, our call, what would prove it wrong, and the date it settles. One card, one URL you can drop straight into a group chat or a memo.
6. Payout day: Public settlement on the due date. Right or wrong, it stays on the record.

What happens when we're wrong

Most publications let their mistakes quietly disappear into the archive. We do the opposite: every call that has an expiration date gets its payout day and what would prove it wrong written down on the day we publish it — and it settles by force when that date arrives.

Here's the ledger right now, and you can count it yourself: 49 calls tracked, 20 already public as verdict cards, first payout date September 1, 2026, and 36 of those already pushed to a public GitHub mirror (bilingual, each carrying its source and payout date) carrying a third-party timestamp we can't go back and edit.

Nothing has come due yet, so we genuinely have no hit rate to show you — an empty scoreboard beats a fake one. Starting in September this table starts keeping score, and a call we lose doesn't quietly get removed.

See the verdict scoreboard →20 verdict cards already public, each one carrying a payout date and a reversal condition. The first one comes due 2026-09-01.

What's on the site, and who each part is for

Four sections, split by the judgment you need — not by topic. The daily brief hands you today's evidence, deep dives take one position to the bottom, Monday's column audits what other people are saying, and verdict cards settle what we said ourselves. Read any one alone; follow all four and you also hold something rare — a public record of which calls held up and which got overturned.

Daily brief: New evidence from the last day or two that changes decisions — including what we chose to skip. For anyone making daily calls on the AI industry.
Deep dives: One judgment, dug to the bottom: evidence, an adversarial table, and what would prove us wrong. For readers who'll spend twenty minutes to truly get one thing.
Monday's Claims & Receipts: The week's most-traveled named claim, checked against the receipts cell by cell — supported, contested, or refuted. For anyone who wants to know if it stands.
Verdict cards: What we said, settled in public when it comes due — misses kept next to hits. For anyone who needs one link to end an argument.

What we haven't done yet

Honesty is the only pitch this system can actually defend, so the gaps get listed too. Every line below is followed by what it means for how you should read us right now:

Receipt discipline currently runs on generation-time rules — no separate program re-scans every citation after the fact. → Citation is our discipline, not our proof. If you want to be sure, click through to the primary source and check it yourself.
A large share of our fact layer still rests on a single source. → That's exactly why those items stay facts, not calls. Anything promoted to a call has already cleared the two-source bar.
The scoreboard hasn't settled a single call yet — zero track record. → Judge us right now on whether this process is legible and whether it actually blocks things. Judge us on results starting in September.
We can't get scoops, and we never will. → Nobody here is calling suppliers for a heads-up. We're a second source — the best verifier and connector on top of the public record, not the fastest leaker.

How to check us

Click the source behind any call. What's linked after each conclusion is the original document, not our paraphrase.
Open a verdict card and look at its payout date. How firm the call is, and what would reverse it, fits on one page.
Check the timestamps on the public mirror. The date a call was written down is a matter of record — not something we get to decide after the fact.

You don't have to trust us — you can audit the receipts yourself.