The calls, graded
A call is a dated, falsifiable statement with our probability attached and the command that decides it. On the grading date the command is run on the fresh export, the result is pasted in, and the tag flips. Nothing is edited afterwards; an ambiguous rule is graded against us. The full text of every call is on the calls page.
| call | the claim | we said | graded on | result |
|---|
The settled weeks
The weekly scorecard, kept to the weeks that have settled entries. The favorite rule buys Yes the first time a bracket's midpoint sits between 65c and 90c, at the ask that was resting plus Kalshi's taker fee, and holds to settlement. "Day before" is the day-ahead entry; "same day" the intraday one. Per contract is the mean over settled entries, in cents, and a week with under 20 settled entries cannot decide anything.
| week of | entered | entries | settled | paid | mean ask | per contract | median size at ask |
|---|
The dead brackets: the ones the station's confirmed running extreme had cleared by a full degree when the robot read them. The last column is the observation feed's error rate, live: brackets called dead that settled Yes anyway.
| week of | dead brackets | book reads | median at 1c | reads with a stale bid | settled Yes anyway |
|---|
the dead-bracket record begins 2026-09-16; the live tables with every week, settled or not, are on the scorecard.
The desk that started this
From July 24 to September 8, 2026, this site was a forecast desk: public National Weather Service guidance turned into a probability for every Kalshi temperature contract, compared with the price, and published when they disagreed. Every published number was scored after settlement against the same official climate report the market settles on, head to head with the price on the same markets.
The money view makes the same point in cents. One contract on every published signal, the first fire per settled market, at the execution price recovered from the fire's own logged quote, held to settlement:
| strategy | total | per contract | 95% day-clustered | win rate |
|---|---|---|---|---|
| Trade the signal (taker: Yes at the ask, No at the bid, plus the fee) | −$147.57 | −4.02c | −5.14c to −3.00c | 42.4% |
| Fade the signal (the other side of the same spread) | −$97.44 | −2.66c | −3.74c to −1.54c | 57.3% |
| Rest on the signal's side at the far edge of the spread, no fee, always filled (maker best case) | +$34.17 | +0.93c | −0.19c to +2.01c | 42.4% |
3,669 signals. The desk's own expected edge on them averaged +20.4c a contract; the realized result is the round-trip friction, a median half-spread of 1c plus the fee. The direction added nothing, on either side.
The desk was closed on September 17, 2026. What survived is the measuring: the logger that had been reading the ladders with no model kept reading them, and its first live week is where the scorecard begins.
| the first live week, 11–15 September, at the logged ask plus the fee | markets | per contract | 95% CI | median contracts at the ask | a day at the ask |
|---|---|---|---|---|---|
| 65–90c the day before (the favorite rule) | 26 | −4.6c | [−26.6, +13.8] | 17 | $90 |
| 65–90c the same day | 193 | +1.2c | [−1.9, +5.4] | 25 | $1,144 |
| 80–90c the same day | 118 | +3.1c | [−0.9, +7.1] | 33 | $977 |
| 90–95c the same day | 106 | +1.8c | [−0.3, +3.8] | 46 | $2,018 |
47,857 reads on 1,200 markets over five days, every bracket of all forty ladders. Five days cannot decide a sign; the size decides the worth, and the day-ahead rule was closed by capacity at $90 a day.
Where the numbers come from
The calls are graded by the commands they name, on an export of the database (agents/export_ledger.py). The settled weeks are scorecard_views.sql, a materialized view over the ladder logger's rows and the outcomes, refreshed every fifteen minutes. The retired desk's record is the closing research note in the repository, RESEARCH_2026-09-10.md, every table of which names the script that produced it and can be re-run on the archived export: desk-versus-market Brier on identical reads with day-clustered bootstrap intervals (agents/replay.py), the money view (agents/exp_pnl.py), and the favorites at the real ask (agents/exp_favorites.py).