Brief
Run the starter harness over chunk {256, 512} x overlap {10%, 20%} with two embedding dims; publish the P@K / Recall / MRR table and defend one winner.
Dataset
DS program RAG corpus + fixed query set + gold-chunk labels (starter bundle). Same corpus reused by m2-l12 eval set; do not retune on the eval queries.
Starter guardrails (plan 88)
Use the provided harness; change only chunking config + embedding choice; log cost/latency per cell; no eval-query peeking.
100-pt rubric
Table completeness (all cells x all metrics) 30; winner justification 25; cost/latency reporting 20; reproducibility (config + seeds logged) 25.
Deterministic checks -> LLM review
Checks: 4+ cells reported; P@K/R/MRR numeric in [0,1]; config file present. LLM review grades the justification and staleness/conflict handling notes.