There is a version of Estelle that is more impressive than the one we ship. We know it well because we keep sketching it and then talking ourselves out of it.
In that version nothing ever abstains. Every question gets an answer. Numbers appear without the paragraph underneath explaining what they cover. The agent reports that the fix is done instead of reporting that the fix is proposed and here is specifically what got checked. It is a much better half hour. It is a considerably worse year.
Being wrong is survivable, being indistinguishable isn’t
Trust doesn’t die because a tool was wrong. Everyone has shipped bugs; nobody expects software to be perfect.
It dies when a tool is wrong in exactly the same voice it uses when it is right. After that happens once you start verifying everything, which means the tool has stopped saving you time while continuing to bill you. The error was not the failure. The failure was that you could not tell it apart from the successes.
So we spend a lot of effort on things that do not demo. Saying which half a check covers. Putting the limit next to the result instead of in a footnote. Marking the difference between a sample and your live system. Refusing out loud when there is no evidence.
The same rule in the other direction
This is not only about avoiding overclaims, and the reverse mistake is the one we actually make more often.
A note saying a capability is not ready, left up after it shipped, is a false statement that feels responsible. We have done this. We left caution in place long after it stopped being true, and caution is harder to catch than hype because it reads as integrity and nobody audits integrity.
The standard is not modesty. The standard is that the description moves on the day the fact moves, whichever way it moved.
What we are actually selling
Not the fastest agent. A floor, and a system that tells you plainly when you have walked off it.
Harder to market. Much easier to keep using.