autoresearch.md: bump current best to 425 samples (run #80)

This commit is contained in:
yusufcanislek
2026-04-26 15:40:01 +03:00
parent 9f539d46f8
commit 358f3e1e65
+2 -2
View File
@@ -63,8 +63,8 @@
- notes: harness initially printed METRIC lines but `run_experiment` did not capture stdout. Replaced the powershell+heredoc body with a powershell wrapper that resolves py.exe and runs an inline python `-c` script; stdout is still empty under run_experiment on this host but `bash autoresearch.sh` from `proxy_bash` reports the correct metric, so I log keeps with `force: true` after manual verification.
## Current best
- metric: 424 vm-shaped samples (run #79, commit b68e779)
- why it won: 139 cumulative new samples across 79 logged runs in this segment. Twenty-one sibling-op pairs closed; latest: extending fnv1<>fnv1a op-order pair from byte to word lane (vm_fnv1_word64_loop pairs with vm_fnv1a_word64_loop).
- metric: 425 vm-shaped samples (run #80, commit 9f539d4)
- why it won: 140 cumulative new samples across 80 logged runs in this segment. FNV-1 now spans byte/word/dword strides paralleling FNV-1a coverage; latest run added vm_fnv1_dword64_loop completing the op-order pair lever at u32 stride.
## What's Been Tried
- experiment: vm_callret_loop with explicit return-PC stack (rstack[rsp])