cribl-decision-2-mini (Cribl)v1.7 · 10 s rule
Data: JevBench v1.8.1, published 2026-10-30 · method v1.7 (10 s rule), measured under the rule
Capability 53.2 · Not ranked in the current release
Measured on the 27 Oct 2026 draw, which is not yet equated to the reference scale; stronger anchor runs for that draw are being measured. Shown unranked until then.
Last measured: 2026-10-29 · measurement release: v1.8.1
Apache-2.0
intelligence
21.5
calibration
84.9
speed
93.1
cost
89.4
Composite (secondary)
9.2 · Not ranked on this board
Cost and measurement conditions
$0.0022 per 1,000 decisions (estimated)
GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 0.8B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 553,650 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.0022479286593751/1,000; old list-price estimate USD 0.00371826729348556/1,000 (deepinfra:Qwen/Qwen3.5-0.8B at USD 0.01/M input and USD 0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0022479286593751/1,000. Estimated, not a bill. USD 0.00224792866 per 1,000 decisions on the frozen 1,489-item common set of this draw (USD 0.00224160455 over all 1,500 items).
p50 latency: 0.2 s
Self-hosted latency estimate under the shared method (assumption, not measured)
Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition (97,887 MiB reported, driver 580.178.04), FIN-03 (Finland, EU), instance type 1RTXPRO6000.30V, image 24.04.cuda13.0. Measured 29 Oct 2026. One pod for the two rows in sequence, with a shred and fresh network cut between rows. First-party loopback System One wire around the unmodified author decision library, pinned container and local hash-verified model, read-only mounts, private output, empty credentials and offline flags. Both rows passed actual-model public smoke before either sealed read. Public successful timed probes at 256/1,024/1,714 tokens passed with native/offline usage equality, separate warm-ups and distinct timed bodies. Public typed smoke covered noul, score, duplicate score levels, choice, nine-option menus, Unicode JSON, special-token text, long choice and long score; native/shim distributions matched. The 80,476-token public refusal probe returned HTTP 422 on both fresh timed repetitions. Fit required 1,500 s for each row; pair-fit required 3,600 s against 5,384.768 s remaining. Forecast over 10 s was 0 for each row; this was an admission estimate, not a scored observation or exclusion. The scored server passed a separate public pre-warm. Runtime and tokenizer hashes were retained; post-sealed health was ready at the pinned revision with max_input_tokens 4,096. Complete 1,500-row run; scored loop 47.434959 s. Raw output sha256 b863a81ff82150671a7217d7c9c76b1220b9d8a7eeadb37df10cfd1ba73ca743 (independently verified sealed artifact). Completion receipts confirm both rows finished on one VM and that the exact VM was closed. Safe-shred receipts confirm the measurement copies were shredded and the retained artifact checksums were unchanged. GPU rental is separate from the benchmark's estimated model tariff.
Measured latency over 1,500 decisions: p50 0.029 s · p95 0.042 s
0 decisions over 10 s.
A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.