mirror of
https://github.com/lightningnetwork/lnd.git
synced 2026-08-20 13:27:27 +02:00
5200d1fc07
In this commit, we land the verdict for the two Opus 5 reflection arms and close exp-010. The headline: opus-default produced the first evolved candidate to statistically tie a champion on any tier (split-val +0.005 vs mx_c3, higher raw success), evolving persistent parallel flow plans, concurrency-first dispatch, and residual-aware shard budgeting across 1,931 lines -- then paid for that corridor depth with a collapse on every off-corpus tier (hard test 0.303 vs 0.583). The medium-effort arm is the cautionary tale: best val score of the family, worst held-out -- caught by the sealed sweep exactly as intended. Champions remain hb1 + mx_c3, now validated against three proposer lineages on this corpus. At fixed eval budget, reflection quality beats reflection throughput; proposer strength moves candidates along the specialist-generalist axis rather than lifting the curve. Both winners are archived, exploit-grep clean. |
||
|---|---|---|
| .. | ||
| experiments | ||
| DECISIONS.md | ||
| IDEAS.md | ||
| NOTEBOOK.md | ||