Journal
Notes from the build.
Writing on incident evidence, regulated service delivery, and the engineering decisions behind Signum — the reasoning, not the announcements.
-
Why capture writes to disk before it acknowledges anything
Buffering an incident channel in memory is fine until the process dies mid-incident, which is exactly when the record matters most. Signum appends every event to a write-ahead log and syncs it before returning, so a restart loses nothing beyond the write in flight — and replay quarantines an unreadable record rather than refusing to boot.
-
A claim that cannot be traced should say so
The failure mode of an AI summary is not being wrong; it is being plausible. Every claim Signum extracts carries the identifiers of the messages behind it, and an extraction whose grounding falls below threshold is refused outright rather than written with a caveat nobody reads.
-
Redaction that fails closed
Suppression that breaks open is worse than none at all, because it is trusted. Signum redacts at read time, per consumer, and any internal error in the redactor returns fully masked output instead of the original text — the transcript keeps what was said, and each reader gets only what they are entitled to.
Full entries are being written up. To get them by email as they appear, write to hello@signum.uno.