Working notes
Notes from the frontier.
We publish the method before we sell the number, which means this is where the arguments get made in public — including the ones that go against us. Three pieces are drafted. None are published yet, and we would rather show you the queue than pretend it is empty.
- 01
Backtesting Bedrock against Typhoon Kristine (2024)
We ran our per-address loss estimates against a storm that already happened. This is where the model held, where it did not, and what we changed because of it — published either way, because a backtest you only release when it flatters you is marketing.
- 02
The false-precision problem in expected-loss reporting
An AAL figure with four decimal places implies a confidence the elevation model underneath it does not have — and every disclosure built on it inherits that false confidence. How we cap precision, and why the cap belongs in the output rather than the footnotes.
- 03
Flood segmentation on solar power, in the dark, mid-storm
What actually breaks when you put an AI accelerator, two cameras and a 3D-printed chassis where the flooding is. Thermal behaviour, night performance on the NoIR sensor, and the power budget that decides how long a station survives an outage.
Want one of these early, or want to argue with it before it’s public? Write to us — reviewers are more useful to us than readers.