Say you want to know: will the Fed raise rates next month? If you're a regular person, you Google it. You find three headlines — one says yes, one says no, one says maybe. You pick the one that sounds most convincing, or the one from the source you trust most. That's not a forecast. That's a vibe check. If you ask ChatGPT, it gives you a confident paragraph. "Based on current economic indicators, it appears likely that..." It sounds smart. But it has no track record, no accuracy score, and if you ask the same question tomorrow with slightly different wording, you might get a different answer. It doesn't know how often it's been right before — because nobody's counting. If you check Polymarket, you get a number — say 55%. That's better. Real money is behind it. But it's just a price. There's no explanation for why it's 55% and not 40%. No evidence trail. No reasoning you can walk through. And if the number changes overnight, you have no idea what drove it. ☘️ Here's what happens inside Yarrow. The same question goes to many independent analysts. They're running on three different AI models from two different providers. None of them can see the market price. None of them can see each other's work. Each one gets the same frozen evidence pack — official data, dated sources, verified statistics. No live search. No Twitter. No vibes. They each have to produce a complete chain: what mechanism drives the outcome → what the data says → what probability they assign. Before any of that ships, the evidence pack runs through five checks: are the numbers internally consistent? Is anything missing that should be there? Is there any information that could only be known after the fact? Has an independent red team tried to break the argument? The final number is the median of the valid analysts — not a vote, not a discussion, not a compromise. Just the middle of six independent judgments. Then it gets frozen. Locked in git. Published. And when the outcome is known, scored — right next to every other prediction we've ever made. Google gives you headlines. ChatGPT gives you confidence. Polymarket gives you a price. Yarrow gives you a probability with the evidence behind it, the reasoning you can audit, and a track record you can check. That's the difference between looking for an answer and building one.

On June 23, 2016, the UK held a referendum: remain in the EU, or leave. Betting markets gave Remain an 85–90% probability. Pollsters predicted Remain. The UK government had no contingency plan for Brexit, and the prime minister had effectively staked his political career on an outcome he himself could not predict with certainty. Brexit won. 51.9% to 48.1%. The pound plunged 10% overnight, its largest single-day drop since 1985. David Cameron resigned the following morning. Global markets lost $2 trillion in value within two days. And Britain spent the next four years trying to figure out what “Brexit” actually meant. The data telling a different story was already there. Online polls had consistently been closer to the truth than telephone polls. The turnout models had never been properly validated against a nationwide referendum. And anti-immigration sentiment in northern England had been systematically underestimated. Betting odds were anchored to the same telephone polls that were later shown to be wrong. No one stress-tested the consensus. An 85% probability became a “fact,” and every plan was built around it. ☘️ Now imagine if the UK government had used a system like Yarrow before the vote. The system receives one question: “Will the UK vote to leave the EU?” Yarrow has independent analysts approach the question from different bodies of evidence: polling data using two different methodologies, regional turnout models, historical precedents from referendums versus elections, and sentiment data from regions that conventional polling struggled to reach. Three analysts flag online polls showing Leave ahead. Two build turnout scenarios in which older, non-urban voters who leaned toward Leave turn out at higher rates than assumed by the polling models. One points out that betting markets are pricing the same telephone polls—not independent information. The system ultimately returns: Remain: 58%, Leave: 42% And it adds: The 85% consensus probability for Remain depends heavily on a telephone-polling methodology that had never been validated in a nationwide referendum. Cameron might still have held the referendum.But the Treasury might have had contingency plans ready. The Bank of England could have positioned itself in advance. And perhaps $2 trillion would not have vanished from global markets within 48 hours simply because everyone was blindsided by an outcome that, if someone had looked at the full evidence base, was clearly possible. The most expensive predictions aren't the ones that turn out to be wrong. They're the ones that are wrong—and that nobody prepared for.

What should a prediction actually look like? Most AI tools give you a number. "There's a 70% chance this happens." Maybe a paragraph of explanation. That's it. No source trail, no what-if analysis, no way to check later whether 70% was the right call. 👀 Here's what we think a prediction should contain — at minimum — before anyone should take it seriously: 1️⃣ a probability that means what it says. When the system outputs 70%, events like that should happen about 70% of the time. Not "I'm pretty sure" dressed up as a number. Actual calibration, measured over hundreds of questions. 2️⃣ a reasoning chain you can walk through. Not "based on our analysis" — the actual evidence, the actual logic, step by step. Which data points support the conclusion? Which ones cut against it? If you can't trace from evidence to number, it's not a forecast — it's a vibe. 3️⃣ the conditions under which the prediction breaks. Every good forecast comes with its own kill switch. "We'd revise if X happens." "This falls apart if Y turns out to be true." If a prediction doesn't tell you what would prove it wrong, it's not brave — it's unfalsifiable. 4️⃣ a frozen record before the outcome. Anyone can claim they predicted something after the fact. The test is whether the prediction was locked in — with all its reasoning — before the answer was known. No edits, no revisions, no "well what I actually meant was." 5️⃣ a public track record. Not just the wins. The misses too. Side by side. Scored by the same rules. Because a system that only shows you its highlights isn't building trust — it's building a marketing deck. ☘️ This is the standard Yarrow holds itself to. Every prediction frozen in git. Every outcome scored. Every miss published alongside every hit. Not because we're always right. But because a prediction you can't audit is just an opinion with better formatting.

We actually tested how Jev performs in a prediction workflow. Yarrow used 48 real-world cases across 31 independent documents. Total model cost: less than $0.10. Yes, we have a cost advantage too. ☘️ Here’s what we found — what worked, what didn’t, and what remains inconclusive. Where Jev did well: Jev can read complex corporate announcements and turn them into structured, verifiable event records. 48/48 cases — it correctly identified the current status. It also caught distinctions that simple keyword matching would miss. For example, one announcement was titled “Withdrawal of Bank Regulatory Application”, while the body also said the company “reaffirmed its full-year guidance.” A keyword-based system could mistakenly interpret this as a withdrawal of the guidance. Jev correctly distinguished the two. Where it failed: Understanding the current status is not the same as determining whether a contract has already resolved. In 10 of 48 cases, Jev correctly recognized that the available material did not mention the outcome — and then incorrectly concluded that the outcome was “No.” For anyone trying to use it in production, this is a critical issue. Overall judgment accuracy:Earnings guidance: 22/25 HBM supply: 16/23 Neither cleared our 90% threshold. Stability is still an issue. We ran the exact same inputs repeatedly. In 8 out of 48 cases, at least one field changed. We also changed only the order of the input paragraphs. In 20 out of 48 cases, the result changed. That means it is not yet suitable for automated resolution or fully unattended workflows. Can Jev predict? We’ve already frozen 12 forward-looking questions — covering Fed meetings, weather dates, and geopolitical events — along with 6 corporate disclosure predictions. The outcomes are still pending. But there is already one signal worth investigating: For companies with regular quarterly disclosure patterns, Jev assigned 93–99% probability to the statement “the disclosure has not been released.” We’re watching this closely. There’s another interesting finding: When we asked about the same event in two different formats — binary Yes/No versus multiple-choice — the probabilities shifted by an average of 12–13 percentage points. The probability depends on how you ask the question. Our interim conclusion Jev is genuinely useful for one thing: Turning messy corporate announcements into structured, checkable records. It’s fast. It’s cheap. And it reads surprisingly well. But reading accurately is not the same as judging accurately. Understanding is not prediction. We’ll keep experimenting and keep testing. And if you’re currently researching Jev — especially its applications in prediction — we’d be happy to compare notes and discuss the results together. 👀 ☘️