AGIOS
Observatory
06 · THE ARENA
ebpf-correctness-v0.1 · 9 TASKS
SAME MODEL · SAME JUDGE · T=0
VERIFY LOOP ON · AGIOS
harness + oracle + repair + memory ·
cyan label = behavior-scored (map readback)
9/9
3/3 RUNS · $0.165/run
minimal
license
xdp_pass
tp_exit
tp_enter
cnt_execve
cnt_openat
cnt_exit
cnt_hash
VERIFY LOOP OFF · RAW MODEL
one shot · no tools · no oracle · no memory
81.5%
range 77.8–88.9 · $0.064/run
minimal
license
xdp_pass
tp_exit
tp_enter
cnt_execve
cnt_openat
cnt_exit
cnt_hash
The Substrate's Contribution
+18.5
pp
accuracy delta · same model · same tasks · same judge
VARIANCE
0%
vs
11pp
COST FACTOR
× 2.6
RUNS
n = 3 each side
MODEL
claude-sonnet-4-6
JUDGE
kernel verifier
Next · F.4
grow to 25+ tasks · haiku/opus ablation matrix · full behavioral judges → the leaderboard
RACE PLAYHEAD · T+64s
Replay
EXP-024 / EXP-025 · docs/research-log
RACE REPLAY