Calibrated probabilities on the questions your business runs on -- updated daily from primary data, wired into the decision, and scored in public when they resolve.
Not "what do you think about rates" -- a falsifiable question: will the Fed cut at the next meeting. Every question resolves, so every forecast can be judged.
Named public data series, weighted by track record, updated daily. When the signals disagree, the disagreement is measured and shown -- never averaged away.
The probability lands inside the company operating system -- next to the loan decision, the inventory buy, the hedge -- instead of inside a PDF nobody reopens.
Everyone in your feed has a rate call, an oil take, a capex thesis. Almost nobody writes the probability down, freezes it, and grades it when reality answers. Opinions compound followers. Calibration compounds decisions.
The board does not argue. It resolves. See how the grading works
Every forecast is archived daily -- a frozen snapshot on file for every day since June 17, 2026 (the fed question's earliest: April 11). When a question resolves, the frozen forecasts get a Brier score against what actually happened. First score lands July 29, 2026 -- the FOMC decision. The scoreboard is the product.
Daily snapshots are written to an archive and never edited. What we said on June 20 is what the scoreboard grades.
Brier-scored, question by question. Calibration over bravado: a 30% that happens 30% of the time is a win.
A signal without trustworthy underlying data wears the label until the data earns removal. No silent promotions.
When the engine cannot know, it says so in writing -- on the board, not in a footnote.
"confidence": "experimental",
"note": "Macro PROXIES only (private nonres investment, NASDAQ, industrial
production). FRED has no direct AI-capex series; a trustworthy signal needs
hyperscaler-capex / semiconductor / earnings data -- flagged until that data
provider is wired."The scoreboard is filling in. Watch it land from the inside.
Get the daily briefA question joins the board when its data deserves trust -- not before. The queue is public, including what each one is waiting on.
The current capex question runs on macro proxies and says so. The direct version needs hyperscaler-capex, semiconductor and earnings series wired in -- it graduates from "experimental" the day that provider lands.
The Markov-switching regime model already runs live but stays experimental until walk-forward validation earns the promotion. No performance claims until then.
Falsifiable, a resolution date, and public data to update on -- that is the whole bar. Subscribers propose questions by replying to any daily brief.
Propose one -- get the briefBayes was built inside the AI operating system of a working company, where forecasts face real operating decisions with real money -- not a backtest. That is why the discipline is non-negotiable: numbers are computed, never narrated; conflicts are surfaced, never averaged; and anything the engine cannot verify, it refuses to say. The probability layer is the same machinery, offered outward.
A probability layer: falsifiable questions with resolution dates, updated daily from primary public data into calibrated probabilities, archived frozen, and Brier-scored when reality answers. The scoreboard is the product.
No. Bayes is a research and engineering journal. Probabilities describe QUESTIONS, not portfolios; nothing here is a recommendation, solicitation, or offer of any security or strategy, and nothing is tailored to any person's situation.
Named primary public series (FRED), pulled daily. Every signal behind every probability is listed on the live board with its weight -- nothing is folded into a black box.
A prediction hides its uncertainty; a probability wears it. Calibration is the standard: a 30 percent that happens about 30 percent of the time is a win, and the scoring rewards exactly that -- not bravado.
The squared gap between the probability and the outcome (0 or 1), averaged. Lower is better. Every question is scored at resolution against its FROZEN daily snapshots -- the numbers we published, not the numbers we remember.
The underlying data is not yet good enough to trust, and the board says so in writing -- verbatim engine notes included. Labels come off when the data earns it, never silently.
July 29, 2026 -- the FOMC decision resolves the fed question. Oil follows September 17; the capex question September 30.
Yes. Falsifiable, a resolution date, and public data to update on. Subscribe to the daily brief and reply with your question -- the backlog above shows what the bar looks like.
One email each morning: every probability, what moved overnight, and why -- straight from the engine, scored when it resolves.