Capstone: the source-grounded research assistant

Conclusion

You shipped a source-grounded research assistant: it answers from a versioned corpus, names the documents behind each claim, refuses when nothing supports an answer, and keeps a trace of which of the two it did.

What you did

• Nothing in it was written for the capstone — every part arrived as one session's deliverable, with its own tests, and the project is the wiring.
• The integration rule held that together: wiring, never logic, so a failing run has one layer to look at instead of six.
• cap01-e1 read a cited answer, cap01-e3 read the trace kinds, cap01-e4 read a green eval gate over the golden set.
• cap01-e5 is the one you wrote, and it is the list that makes session 14 a backlog rather than a blank page.

The failure you handled

A question the corpus does not answer. Retrieval comes back empty, the run refuses before a model call is spent, and cap01-e2 reads the empty calls list as the proof.

The second one behind it: a citation retrieval never returned, stripped rather than printed, with the flag and the trace line that say so.

What to carry forward

That list is also demo day's honest limitation, prioritised weeks before anybody asked for it.

Into session 15

Session 15 assumes the traces are on, the eval report is a number you produced, and you can say which architecture decision you would defend and what would reverse it.

It does not assume the demo goes well: you draw a failure you did not prepare for, and failing safely with an accurate sentence is a pass. The one answer that fails is claiming it worked when the trace says otherwise.