Skip to content

How we know what we know

Every claim on the reference pages carries a label, and the labels are less a rating than a record of how we got there. This page explains what each one means and tells the stories behind them, because the stories are where the rules came from.

Verified means we checked the claim against a file or bytes. Confirmed independently means a second, separate route landed on the same answer, such as a firmware image that names a version and a live drone that reports the same version over its normal protocol. Agent-cited means an automated reading was spot-checked but not fully re-derived. Inference is our own reasoning from something verified, offered as reasoning. Open means not established, and we would sooner say that than fill the space. The evidence ledger gives every claim its status and where the evidence lives, and a few rows also carry a Corrected status for earlier statements that turned out wrong.

The ledger says that all its rows were read through together on 3 October 2026 for consistency. The next sentence says that this checks that rows agree and is not a fresh check of every claim against the files. Those are different things. Rows can agree perfectly and all be wrong, because agreeing is what happens when they share a source or a belief.

We learned the sharper version of this from the simulator. We built a protocol-accurate fake drone so the client could be tested without the aircraft, and 362 tests passed. The first live run found three protocol bugs within minutes. The simulator and the client had been written from the same assumptions, so they agreed, and agreement proved only that the code matched its own beliefs. A test double built by the same mind as the code is a consistency check and not a correctness check.

The flight state machine has 37 small diagrams, drawn from a table by a script. The sources parsed with the real Mermaid parser, and that check was passed before anyone had looked at a picture. It said nothing about whether the pictures could be read. Only rendering them, on a phone as well as a desktop, showed the first versions were far too wide, and they were redrawn. The flight state machine page records this at the end of its diagrams section. A parser answers whether the text is valid. Whether a person can read the result is a separate question, and only looking answers it.

Almost all the firmware work was reading: files, bytes, configuration lines and decompiled code. Nothing executed any firmware. That is strong for some questions, such as what an image contains, and quiet on others, such as what a running unit really does. The clearest example is which configuration set the drone loads. Nothing in the image says, and the answer came from one unit’s own log. We hold it as confirmed for that one unit and open for the rest.

The same split applies to the phone app. Its code says what it can do, and the drone’s log says what it did.

Absence is slippery. A probe of the video stream only read the outer header byte of each packet, saw no keyframe type, and the summary became “the stream has no keyframes”. A fragmented packet hides what is inside it, so the probe could not have seen them. “None observed” and “none exist” are different claims. We try to ask what an instrument could not have seen before a measurement becomes a note, and the lesson about the log search for setting writes has the same shape.

Some ledger rows end with a short tag. An id gives a row a name that other rows can rely on. A basis lists the ids a row depends on. The point is practical: when a row changes, you search for its id and find everything that needs rechecking. A row that rests on one file read says basis none, so that “never looked for dependencies” and “looked, found none” read differently. The build checks that every reference resolves. It cannot check that a row names the right dependencies, so a basis is a statement someone made, not a guarantee.

A pass that has never been seen to fail only tells you about the check’s silence, so before trusting a negative result we try to make it positive once.

Comments

Spotted a mistake, or tried this on your own Bebop? Leave a note. Comments are kept on this site's own server.