Aelyx

Asserted knowledge — written by a person and anchored to a commit — is kept separate from derived knowledge that was computed. Collapsing the two is how a project-intelligence layer starts confidently repeating its own inferences back as facts.

It also compares the architecture a project says it has against the one its code actually has, which is only a meaningful question if those two things are stored separately in the first place.

When the context a question needs exceeds the budget available, Aelyx refuses, with an exit code distinct from both success and failure. It does not silently answer from less than it needed, because a degraded answer with the qualifications removed reads exactly like a good one.

What it does return is nested by trust zone, with repository excerpts fenced and marked untrusted and never edited — an excerpt that had been cleaned up would no longer be evidence of anything. It is reached over the Model Context Protocol, where only one tool discloses project content and only one can run an external program.

That is the behaviour the Red Sentra team values in it, and the reason it is used internally: a tool that degrades quietly under load is worse than one that stops.