questions being tested · 3 open · 3 answered · a "no" counts as an answer
open · adk-learning · a week inside Google ADK, having already built the roles-and-mentor version by hand
Six of seven days in: single agent, pipeline, delegation, RAG grounding and cross-session memory all built and tested; the Cloud Run deploy is scripted but its live eval run sits blocked behind free-tier throttling. The framework still decides less than expected. The capstone, and the verdict, remain open.
answered · payments-rag · recall stuck at 0.60
0.60 → 0.60
No lift at all. Big-corpus results from the literature did not transfer. Vector stayed the default.
answered · payments-rag · the fashionable next move
0.60 → 0.70, ceiling
Only the ones already fetched. It promoted every reachable page and then hit a hard ceiling, at seconds per query. Built, measured, left out of the live path.
answered · payments-rag · two evals disagreeing
6/10 → 9/10
The metric. It checked for one labelled page while the answer sat on a different, equally correct one. Score the fact instead of the page and the two evals reconcile.
open · payments-rag · open to anyone, no sign-in
Live behind rate limits and a hard daily cap. Nobody has found the ceiling yet, which is itself the uninteresting half of the answer.
open · payments-rag · every eval rests on it
Ten questions is a noisy ruler, and picking more of them is not obviously the fix. What makes a question worth including is the open part.