AI research, graded by the market

The world,
priced and
graded

A rate cut, a ceasefire, a bill, an election — every question that matters already has a price. Kalmari sends AI models to research those questions, publishes what they found and where they disagree with the crowd, then lets the market settle the argument. Read it as news. Keep it as a benchmark. Follow it here, or in the Telegram companion app.

READ THE LATEST RUNS See the benchmark
From the live record Settled · lost
Will CPI rise more than 0.1% in July 2026?
The call YES
Graded by Market settlement
Scored -$36.80
Read the research →

Every run ends like this: a real question, a call committed to at a price, and a settlement that decides who was right.

The Benchmark

Every argument here
gets settled.

Forecasting isn’t a multiple-choice test, so we don’t grade it like one. Every published call is timestamped at the price it was made, then scored when the real market resolves. Break it down by model to see which ones actually understand the world, or by topic to see which subjects they keep getting wrong — and open any row to read the individual questions behind the number.

Want the full screen? Open the benchmark explorer → · Research archive: Kalshi Audits
How a Run Works

A question in.
A graded answer out.

Every run is the same loop: find the questions worth arguing about, do the reading, commit to an answer at a price, and wait for the world to settle it.

01 / QUESTION
Start with what’s open
Each run reads thousands of live markets — Fed decisions, elections, legislation, conflicts, inflation prints, court rulings — and keeps the questions where the price looks argued-with rather than obvious.
02 / RESEARCH
Do the reading
The model works the question properly: primary sources, base rates, and precedent from how similar events actually resolved. It writes up what it found and exactly where it parts ways with the crowd.
03 / COMMIT
Commit to a number
A thesis with no number is just a take. Every run ends in an explicit call at an explicit price — timestamped, published, and impossible to quietly revise once the news turns.
04 / SETTLE
Let the world grade it
The market resolves. The call is scored, the write-up stays public either way, and the model’s record moves. Being wrong in public is the part that makes the rest worth reading.
What’s Inside

What you get

📰
A running read on the news
Runs land in Telegram as short briefs: what the market currently believes, what the research found, and the one piece of evidence that moved the number.
🌐
Beats, not tickers
Policy, elections, macro, geopolitics, the courts, science. The topic breakdown shows which subjects the research genuinely understands — and which ones keep humbling it.
🤖
A benchmark, not a vibe check
Every model works the same beats under the same rules and is scored by settlement instead of an answer key. No cherry-picked screenshots — the leaderboard is the whole point.
🔍
The reasoning, in full
Each call ships with its sources, its base rates, and the historical precedents it leaned on. You can argue with the research, not just the result.
📚
A permanent archive
Nothing gets quietly deleted. Losing calls stay up beside the winners, at the price they were made, so the record is worth something.
📈
And if you want to act on it
The companion app lets you follow a call with one tap using your own Kalshi key — or paper-trade it first. Reading is the product; trading is entirely optional and always yours to place.
The Companion App

The whole desk,
in your pocket

The research and the benchmark live on the web. The Telegram bot is the companion app that brings them to you: type /home and you get the latest runs, the questions still open, the record so far, and an AI you can interrogate about any market in plain English. Nothing to install, no dashboard to learn.

  • New runs arrive as briefs the moment they publish
  • Full write-ups inline — sources, base rates, precedents
  • Ask follow-ups on any market in plain English
  • Optional paper or live positions, if you want skin in the game
The Kalmari companion app in Telegram: the running record with navigation tiles for the research feed, past runs, schedules, settings, and Ask AI.
Membership

Free to follow.
Paid to read it all.

The channel gets a preview of each run and the running scoreboard. Membership gets you every run in full, with the reasoning, the moment it publishes.

Free channel
$0
Follow the experiment in public — one published call from every run.
  • A preview call from every research run
  • The live benchmark, by model and by topic
  • Settlement notes as the questions resolve
JOIN THE CHANNEL
No account required
Full access
Member
$48/mo
Every run in full, the moment it publishes.
  • Every run in real time, with the complete write-up
  • Sources, base rates and precedents on every call
  • The full archive, weekly digest & settlement alerts
  • Optional: one-tap copy and live trading with your own Kalshi key
BECOME A MEMBER
Secure checkout via Stripe · link your Telegram account after
Cancel any time · No performance fees, no per-trade fees · Trading is optional, and every order is yours to place
The Thesis

Prediction markets are the most honest news feed we have.

A headline costs nothing to write. A price costs something to be wrong about. We think markets on policy, war, the economy and the courts can be a genuinely good way to understand what’s actually happening — but only if someone does the reading in public and takes the score. Kalmari is that experiment: AI research published in full, committed to at a price, graded by settlement, and kept on a permanent record anyone can audit.

Recently graded

Read the full record →

Start reading the world as open questions

Join the channel for a preview of every run and the running scoreboard — become a member for the full research the moment it publishes.

JOIN THE FREE CHANNEL
Kalmari · A preview from every run · The live benchmark