How this brief is made
This page covers two things: how the daily brief is made, and how to check what we write. SecondSource is an AI system that reads sources and drafts, then runs automated checks on admission, wording and pre-publication quality; the rules and the key calls are decided by a human. We don't ask readers to trust it — the sources, the settlement rules and the corrections are public, so anyone can check.
The name comes from the chip industry, where a second source is a qualified alternative supplier for the same part, so no buyer has to depend on just one. Applied to research, it means depending on no single source, and on no one person's expertise.
Two tracks: the daily brief, and calls that are checked when they come due
The daily brief
- Read primary sourcesFilings, official announcements, regulatory disclosures, papers and practitioner interview transcripts
- Check and labelSourced facts, our analysis and unverified news are labeled separately
- Write three editionsChinese, English and Japanese share the same research and evidence; each is rewritten for its readers
- Publish each dayThe site carries the full text and its sources
The brief goes out the same day; it doesn't wait for any call to come due.
Calls checked when they come due
- Write a verdict cardThe call, what it rests on, what would overturn it, and a scheduled check date, fixed on the day it is written
- PublishEach card has its own URL; most are also copied to a public GitHub mirror
- Check when dueJudged against the rule written in advance; if the named data isn't out yet, the card says what it is waiting for
- Keep the result on the cardHits, misses and calls the evidence can't settle all stay; if a rule was ambiguous, the fix is dated and explained
A result can appear on the scheduled check date at the earliest.
Three kinds of content in the daily brief
- Sourced facts: each one links to the original document, and only what can be quoted from the source goes into the body.
- Our analysis: how we read the event and which part of the industry it moves, marked as our reading.
- Unverified news: kept in its own section, "Also happened — not verified by us yet", with who reported it; never treated as fact.
The brief earns its place when it helps readers see more precisely how technology turns into products and reshapes companies and industries; an item doesn't have to trigger an immediate purchase, investment or other decision.
Insights in the research graph clear a higher bar
Behind the brief we maintain a research graph — a database of facts, insights and theses. Facts clear a low bar: verifiable, in coverage, structurally needed. Insights clear a high one: at least two independent, reliable sources, and a senior decision-maker somewhere in the industry would change a judgment after reading it. Anything resting on a single source stays a hypothesis, with capped confidence.
That bar applies to newly admitted insights; a few insights admitted earlier have not yet been re-checked against it, and pages do not yet mark which ones. Every admit or reject of an insight is logged, blocked and parked ones included; facts are not logged one by one. The ledger is internal for now; what readers can see are the verdict cards and the corrections page.
How forward calls are checked
A call that can expire is written down with its scheduled check date, what counts as right, what counts as wrong, and what would overturn it — all on the day it is made. On the check date it is judged against that rule, and the result stays on the card, hits and misses alike.
A result can arrive late for a few reasons: the named data source hasn't published by the due date, the check itself was done late, or the evidence can't settle it. Each card shows the scheduled check date and the date it was actually settled; a card without a result says what it is waiting for, or that the reason was not recorded. A call past its check date without a result is not a miss.
If a rule turns out to be ambiguous — two conditions overlapping on the same number, for example — the correction appears on the card with its date, the original wording and the reason. The original record is not deleted.
Some calls have already come due. Settled ones carry their result on their own card — right or wrong, it stays on the site, and a call we lose doesn't quietly get removed; ones past their check date without a result say what they are waiting for. We don't post an overall score here; read them card by card.
What the public version history does and doesn't prove
Most verdict cards are also copied to a public mirror on GitHub. Its version history lets anyone see what changed on a card, and in which commit.
Git dates, however, are written by whoever makes the commit, so on their own they are not independent proof of when a call was first published. The first batch of cards was backfilled into the mirror on 2026-08-03, so the mirror's record for them starts on that date.
The cards on this site and the files in the mirror are two lists with different entry rules, so they don't match one to one:
- The mirror takes only calls that came in before they were settled and whose test can be settled as written. Calls whose test turns on interpretation, that can't be settled at all, that are missing a required field, or that haven't yet been checked for whether their test can be settled as written stay out of the mirror, and are still public on this site. Once a call is in the mirror it stays; when it settles, the status line in its file is rewritten with the result and the settlement date.
- Whether a call has its own page here depends on whether it has been in the card list or linked from a daily issue. The list shows one batch at a time: calls past their check date or already settled come first, then calls with an in-depth report, then those with the nearest check dates. A page, once published, stays up. That is why some calls in the mirror don't have their own page here yet.
How to check us
Try it on a real card. We got this one wrong, and the result stays on the card:
- Open the card. The call, the receipt, what would overturn it, and why we marked it a miss, all on one page.
- Click its sources. Each one links to the original document, not our paraphrase.
- Click “Public version history” on the card. It opens the card's version history in our public GitHub mirror, where every change is visible; the link next to it is pinned to the first public version. Git dates are written by whoever commits, so read them as version history, not as independent proof of time.
What we haven't done yet
Honesty is the only pitch this system can actually defend, so the gaps get listed too. Each line is followed by what it means for how to read our output right now:
- Citations are policed by writing-time rules — no separate program re-scans every citation after the fact. → Citation is a discipline, not a proof; to be sure, follow the link to the primary source.
- A large share of the research graph's fact layer still rests on a single source. → That's why those items stay facts. Newly admitted insights must clear the two-source bar; a few admitted earlier have not yet been re-checked, so some may still rest on a single source.
- The public mirror can lag behind the site: the latest results and corrections may not be in it yet. → For results, the verdict card on this site is the record; the mirror is for comparing versions of the call and its test.
- We don't publish an overall score. → Each result is on its own card, misses next to hits; read them card by card.
- No scoops. → Nobody at SecondSource calls suppliers for a heads-up; the calls are built by checking and connecting the public record, not by breaking news first.
No need to trust us — the receipts can be checked.