§ Now / Working Set
The lab's current working set.
A dated snapshot of active experiments, open architecture questions, and the reading shaping the next cycle of work.
Snapshot
§ 01 — Active experiments
- EXP-021
Hybrid retrieval reranking under domain drift
Measuring how reranker quality decays as the underlying corpus shifts week-over-week.
lane / RAG - EXP-020
Controller-worker agent topologies past 6 tools
Comparing single-loop ReAct collapse against an explicit controller split.
lane / Agents - EXP-019
Eval-as-CI: trace contracts as regression gates
Treating evals like CI/CD: trace fixtures, scorers, budgets, and failure ledgers.
lane / Evaluation - EXP-018
Curriculum graphs as capability dependency models
Modeling skill prerequisites as a DAG instead of a flat lesson list.
lane / Learning
§ 02 — Open questions
- Q01Where does single-loop agent reasoning structurally collapse?
- Q02What's the cheapest signal that a retrieval system has silently regressed?
- Q03Can capability ledgers replace traditional curricula for adult technical learning?
- Q04What does observability look like when half the call graph is non-deterministic?