Same grok-4.6 SuperGrok seat. No SPEC.md. n=3. Task pass ≥2/3.
Honest list$ is the official low band ($2 / $0.50 cached / $6 out per 1M)
on passing reps of passing tasks. Failed tasks contribute $0.
SuperGrok cash is $0. Wall is a hang detector. Do not score Graff
first-token (boot ›).
Only graff-195 is G1–G6 certified. The other 11 are published live tasks, not a second gate pass.
| harness | reps | tasks | honest list$ | mean wall |
|---|---|---|---|---|
| graff | 35/36 | 12/12 | $21.48 | 264s |
| grok | 33/36 | 11/12 | $33.69 | 362s |
| OpenCode | 36/36 | 12/12 | $25.71 | 309s |
| Pi | 35/36 | 12/12 | $18.47 | 334s |
| exo | 25/36 | 9/12 | $16.01 | 281s |
A check-green with no tokens does not count. exo's last two turbo
tasks (turbo-ws-duplex, turbo-asgi) exited in <1s with no usage —
the public check still passed on the untouched tree. graff-gemini-ix
is 1 real pass. Grok still drops #727. Stored JSONL list$ high-bands
the rep sum (~2×); do not tweet that column.
Machine-local JSONL (gitignored): run-20260908-173944 (#195),
run-20260908-182322 (graff rest), run-20260909-011118 (grok),
run-20260909-042501 (OpenCode), run-20260909-095017 (Pi),
run-20260909-095016 (exo). Rebuild the card with
python3 graff-evals/plot_live.py --from-jsonl.