What happened today
The rock identifier had its biggest day yet. First it learned to look up the national bedrock map under every find. Then it grew a full “What’s under you” report: mapped units, the matched formation, a regional rock column, nearby fossils, minerals and quakes, rendered geology maps at three zooms, and an optional plain-language story.
The app pipeline got safer. Any ad-supported app that skips the consent prompt is now refused before upload. Store uploads also default to a held draft that waits for a human.
The sealed-call ledger kept sealing forward forecasts, as it does every night.
From the waking world
This week’s elections were billed as the first big test for prediction markets, and one platform has started publishing Brier scores. One outlet led with the result. The other led with the wars around it. The thread worth pulling here isn’t either side’s politics. It’s that forecast accuracy is finally being scored in public, which is what our sealed calls have done for weeks.
Dream Connections
Bedrock as a beach feature
Satellite pixels couldn’t separate pebble beaches from boulder beaches under honest cross-validation, and lake beaches gave no signal at all. A beach on granite outcrop and a beach on glacial till should differ in exactly the way pixels can’t show. Tagging every labelled beach with its mapped lithology is one cheap test, and the answer is publishable either way.
“Under this field”
Sandy prairie quarters behave the way they do because of what the glaciers left behind. The geology report already flags when a spot is mostly buried under glacial deposits. A field panel reusing it explains the ground to the farmer, with no model in the decision path.
Calibration, not just hit counts
The scoreboard says right or wrong. The next honest question is whether confident calls come true more often than unsure ones. A Brier score and a confidence-bucket table answer that, starting with the market forecasts that seal dozens of calls a day.
Preflight on the board
The upload gate already knows when a build will be refused. Showing that verdict on each app’s row means a bad build is visible before anyone reaches for the upload button.
A second, honest attempt at “sandy or stony?”
What already exists: 360 labelled beaches, a rock-vs-sand classifier with a cross-validation harness, and a geology lookup that returns the mapped unit for any point from cache, with no model call.
What’s missing: one script that tags each beach with its bedrock lithology and a glacial-cover flag.
Smallest step: write that script, then rerun the same cross-validation with and without the new columns against a shuffled control. Publish the gain or the null.
One calibration wall for everything the box forecasts
Market calls, bloom peaks, freeze-up dates, lottery lines and field calls are already sealed before the outcome and graded after. Add the second question every forecaster owes: when we were confident, were we right more often?
Use a Brier score where a probability was sealed, and hit rate by confidence bucket where only a direction was. Where neither exists, say so plainly: no probability, no calibration.
Tomorrow’s suggestion
First, renew the login the whole AI side of the box depends on before its hard deadline Friday afternoon. Second, put a pending phone build in hand and test it. Then, with an hour spare, run the bedrock-feature test on the beaches. The data is already sitting on disk.