Capstone: the source-grounded research assistant

Quick quiz (ungraded)

To confirm the model was warm before the question arrived
Nothing in the capstone needs a warm model, and FakeLLM has no warm-up.
Because the refusal alone says nothing about when it happened; an empty calls list is the evidence retrieval decided it
Correct. A refusal after a model call is still a refusal, but it paid for something retrieval already knew for free.
Because a refusal is only valid if the model agreed with it
The deterministic door is the reliable one. Asking the model to confirm an empty retrieval adds a cost and a way to be talked out of it.

Keeps the citation; the model presumably had a reason
The model's reason is unavailable and the claim is unsupported by anything the program can point at. That is the failure the check exists for.
Raises, so the caller sees that something went wrong
The caller gets a traceback instead of a result, and the part of the answer that was usable is lost.
Strips the citation, drops confidence, flags for human review, and records all of it in the trace
Correct. The sentence may still help a person; the source claim beside it may not. Both facts survive.

The chain, and the measured comparison goes in the decision record
Correct. Pick the least autonomy that passes the eval, and write down the number that would reverse the choice.
The reflection loop; it is the more capable architecture
Capability you cannot measure is a claim. Three times the spend for no measured gain is a decision you cannot defend in minute 4.
Both, behind a flag, decided per question at runtime
That adds a branch nobody has evaluated, and the demo now has two systems to explain and one measurement covering neither.

Show the refusal, point at the trace line where it stopped, and say what you would fix
Correct. Failing safely is a pass. A refusal the caller can read plus an accurate next step is the demonstration.
Switch to the recorded run from last night, which worked
That answers a question nobody asked. The room already believes the happy path exists.
Say it works locally and move on
Claiming it worked when the trace says otherwise is the one answer that fails.

The check needs unique keys to store the rows
Nothing is stored. The list is read once and judged.
Two issues at rank 1 is an inventory, not a prioritisation, and the ordering is the whole exercise
Correct. The impact is what orders the list, which is why every row has to carry one.
Ties mean the two issues have the same cause and should be merged
They may be unrelated and equally annoying. You still have to say which one you would fix first.