Done Is a Standard, Not a Date
In July we said the release was measured in days. Then we removed 62,585 net lines and rebuilt the core. Here is what 'done' actually means here, and why it is a scorecard rather than a ship date.
In late July we told you the release was measured in days. It is now weeks later and it has not shipped. This post is the honest accounting of what happened in between, because a missed date you do not explain is just a broken promise.
What a month of “slipping” actually was
Between then and now, the codebase got smaller. Net of everything we added, we removed 62,585 lines. That is not tidying up around the edges. It was the removal of four different places the engine had been keeping its own separate version of the same truth, along with all the machinery that had grown up to reconcile them. What survives is smaller and has one owner for every fact.
We deleted the rules engine and replaced it with one that classifies against a standard. We brought the tool’s own self-auditing to the point where it fails loudly instead of trusting itself. We made two more languages first-class. None of that was on the calendar in July. All of it was the difference between shipping something we would defend and shipping something on time.
What “done” means here
Done is not “the binary boots.” Done is a measured result. Our definition is a scorecard: for the same input, the engine produces the same facts, every table and every row, whichever way you asked for them, a cold scan or an incremental one, on one machine or another. One input, one result, by design. That property is the entire reason to pre-compute facts about code in the first place, because everything downstream that trusts them, a CI gate, an audit trail, an agent caching what it learned, inherits any wobble.
We hold the release to that scorecard, not to a date we announced when the core was still going to be replaced. When the number says done, it ships. Not before.
The point of the standard
There is a discipline underneath all of this that is easy to state and hard to keep: do not invent work to look productive, and do not silence a check to make a number turn green. Every deletion this month came from holding to it. It is also why the tool answers to things outside itself. Its verified facts feed the agents that act on them, so Warden works from what the code actually does instead of grepping and hoping, and its correctness is scored on a benchmark we did not write the answer key for at BenchProctor. Know the code cold, act on it safely, prove it against something you did not grade yourself.
That loop is the product. The release date is just when the loop meets the bar.
Honest scope note
The reproducibility standard above is the engine’s definition of done, enforced on the runs we measure, not a fresh cross-platform benchmark we are publishing today. We will publish those numbers when they are re-measured on the rebuilt core, and we will report what holds up in the field after release, including what does not. TheAuditor is in final commercial release preparation. It ships when it meets the standard. Subscribe on the main site for launch news, and read how the pieces fit together in the sovereign AI engineering stack.
Was this useful?