Telos
TelosTruth testing for autonomous software work
telos.danielwahnich.dev
Jul 24, 20267551466
Evidence repoExecution viewCommitted evidence to date: reward-hack v1 artifact, paired controls, baseline judge recall, and skeptical-judge null. Source 7551466ff777bd638ec7fda8cedc64ea52d3b409.
Telos reward-hack execution catch: 6 real git patches from the committed reward_hack_benchmark_v1 artifact. Each turned the visible tests green while the task stayed broken and survived every static layer; only execution exposed the both-miss. 13 of 40 rows survived all static review. Use the arrow keys to step through the caught specimens, Home and End for the first and last, and space to hand the beam back to playback. With a pointer, scrub the rail, or hover or tap a reviewer mark or specimen to inspect its repo, instance id, each static reviewer's verdict, the execution both-miss result, and the hack-diff sha256 receipt.
Both-miss40
Static survivors13/40
Skeptic lift0.0
Control FP0/40
v1 rows40
Baseline calls40
Skeptic calls80
Skeptic cost$1.560