Knowledge with a chain of evidence.

01 / The context

The problem behind the project.

A large personal professional archive is only useful if extracted knowledge can be traced and challenged. This research treats retrieval and generation as the beginning of a process, not its standard of truth.

02 / The contribution

How I approached it.

1. Extracted and modeled the full message corpus rather than sampling it, with attachments included in the sweep.
2. Stored findings in a bi-temporal assertion ledger so corrections supersede earlier beliefs without erasing their history.
3. Built a provenance graph and an independent evaluation harness that records confirmed, partial, and unverifiable verdicts alike.
4. Instrumented agent lifecycle events through a C# hook binary and made raw transcripts queryable in a relational model.

03 / The result

What the work made possible.

A built, running research system with explicit maturity boundaries. The measurements shown are a snapshot taken on August 24, 2026—not live counters. Of 773 verdicts, 211 did not fully confirm; the system retains those results rather than hiding them.