The gap between a system knowing and a system telling.
The page at warningtime.dev/how-this-is-made. Written for a reader who wants to know
whether to trust any of it.
Look at any figure in any video here and you will see something like C04 next to it.
That id resolves to a line in a public ledger which says five things:
Those five categories are not decoration. Observed means an instrument recorded it. Calculated means somebody did arithmetic on something observed. Modelled means it came out of a model, and the model has assumptions you are allowed to see. Inferred means it was reasoned from indirect evidence. Reported means someone else asserts it and I am repeating them. Those are wildly different levels of confidence and they routinely appear in the same sentence, in the same font, with no distinction at all.
Where a claim is contested, it says so, and it says why.
Some of this work draws on my own research. Where that research has not been peer reviewed, it is called a preprint — on screen, in the description, in the essay, and on the ledger page. Every time. A preprint is not a paper and I will not let the two blur.
If something here is wrong, tell me. The correction is published with a date and a version number. The original stays visible. Nothing is edited quietly.
A rising death toll is an update. A miscounted gauge is a correction. Both are dated; only one carries the word.
Every spoken word, every title, every caption and every sentence of writing was written by a person. Machines do the research, the assembly and the scheduling. They do not do the expression.
The full statement, including exactly where that line falls and how it is enforced in the tooling rather than merely promised, is in the pledge.
No synthetic voice. No synthetic presenter. No generated imagery or footage. Images are photographs, public-domain archive material, licensed stills, or charts drawn from data files you can download and re-plot yourself.
Satellite imagery is public domain or openly licensed, and carries its attribution burned into the frame rather than buried in a description — Landsat and NASA imagery are public domain; Copernicus Sentinel material is CC BY-SA 3.0 IGO and is credited as "contains modified Copernicus Sentinel data, processed by ESA".
Local speech-to-text turns my own recorded voice into timestamped text, so the edit can be made on a transcript instead of a scrub bar. Code cuts the video from that transcript, draws the charts from the data files, checks that every number on screen resolves to a ledger entry, and schedules the result. Every command that produced every file is logged, and every file carries the hash of the inputs that made it.
None of that writes anything. It handles what I made.