decisio v0.9.0 on gemma-4-12B-it (frozen, one prefill per question, self-hosted)v1.7 · 10 s rule
Data: JevBench v1.8.18, published 2026-11-20 · method v1.7 (10 s rule), re-scored from the recorded per-decision times
Capability 70.0 · rank #85 of 251 on the open-weights board
Last measured: 2026-10-07 · measurement release: v1.6.1
Apache-2.0 software; Gemma model terms
intelligence
52.3
calibration
87.6
speed
85.3
cost
67.7
Composite (secondary)
70.1 · Composite rank #82
Cost and measurement conditions
$0.012 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 gemma, 12B active parameters; throughput 26861.9166667 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 779,965 input + 1,477 output tokens over 1,477 decisions = 529.07379824 tokens/decision. GPU-time cost USD 0.0119075160195099/1,000; old list-price cost USD 0.0265536899119838/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.0119075160195099/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases.
Measured p50 latency: 0.144 s (on-GPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 0.44 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.144 s on-GPU, no network (timed on our GPU over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 0.44 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
measured 2026-10-07 21:11:36-21:15:32Z on the live v1.6.0/v1.6.1 pool (inputs/input-selfhosted-1500.jsonl, sha256 901983ae...), 1,477 of 1,500 answered. github.com/aminry/decisio tag v0.9.0 = 333a0a849de7658969a8f8933b34a8097a8c3489 (verified against the GitHub refs API before the run and re-read as the pod's checked-out HEAD after it), Apache-2.0, installed with their own `uv sync --extra serve --frozen` so their lockfile fixes every version (decisio 0.9.0, torch 2.13.0+cu130, vllm 0.30.0, transformers 5.18.0) and served with their own command `python -m decisio.serve.vllm_engine --base gemma-4-12b`. --base carries google/gemma-4-12B-it@707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7 (bf16, 22.8 GiB) and that profile's own settings - temperature 3.592 for every question type, a system turn, the spaced layout read at the chat template's own answer position, every single-token form of each letter summed. Frozen base: no fine-tuning, no adapter, no task registered. UNCHANGED typesafe adapter and UNCHANGED run_v16.py; all eight checks of the unchanged smoke gate passed on synthetic requests before any sealed item was sent. A FOLLOW-UP OF THE EARLIER decisio-gemma-4-12b-v080 ROW (run 54, same pool, same basis, same rate), measured as a CANDIDATE: the two serving versions are reported separately. WHAT v0.9.0 CHANGES is prefix-cache state-boundary registration, request ticketing and engine-health probing; the whole readout package, serve/systemone.py, serve/temperature.py, serve/abstention.py and the entire vllm_plugin are BYTE-IDENTICAL between the two tags, and no temperature or prompt line moved (r61/receipts/DECISIO-V090-SOURCE-DIFF.json). THE 23 LONG ITEMS FAIL WITH A BARE HTTP 500 from the author's own server - vLLM raises VLLMValidationError because the rendered prompt exceeds the 32,768-token context the profile declares, and the server does not translate that into the 422 the harness scores as a refusal. They are counted wrong either way and they are exactly the 23 items the v1.6.1 common cost basis excludes (the ones the Jev reference itself refuses), but it is the author's own code and A FINDING TO REPORT, NOT A THING TO PATCH AROUND; no setting of theirs was changed and no retry was made. One NVIDIA RTX 6000 Ada Generation 48 GB (Lium, Toronto, CANADA) - the same card model and city on which run 54 measured the v0.8.0 row, chosen over the 96 GB class the author names because the 12B profile is proven to fit it. Latency axis uses the standard self-hosted x2 + 0.15 s adjustment. Serial, one decision per request, no retry, no batching, offline after setup, credential-free. add-requests run 61, PREREG-R61-AMENDMENT-2, benchmarkheaven.com/submit ref 65B71BD3 (bh-submit f05c3298-702f-4314-bc8e-f823d9bce858); previously GitHub #169 for v0.8.0
Measured latency over 1,500 decisions: p50 0.144 s · p95 0.264 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.
Breakdowns (not part of the score)
Chance-corrected competence (0 = chance, 100 = perfect) by category on S+P+L1+L2+L3. Raw values, not equated; the same cells feed the compare view.
Capability by subject topic
| Rules, policy & law | 46.7 | n=1616 |
|---|---|---|
| Coding & software | 70.8 | n=382 |
| Math & numbers | 10.4 | n=370 |
| Finance & commerce | 46.8 | n=223 |
| Support & operations | 46.9 | n=160 |
| Everyday language | 74.1 | n=145 |
| Safety & security | 70.3 | n=109 |
Use cases (TypeSafe categories)
| Model routing | 73.5 | n=524 |
|---|---|---|
| Legal & compliance | 52.9 | n=515 |
| Customer support | 49.3 | n=240 |
| E-commerce | 48.0 | n=119 |
| Insurance claims | 58.7 | n=119 |
| Risk assessment | 33.8 | n=118 |
| Financial crime | 22.5 | n=101 |
| Feature extraction | 25.4 | n=99 |
| Lead generation | 48.5 | n=95 |
| Recruiting | 25.8 | n=94 |
| Knowledge graphs | 48.6 | n=94 |
| LLM guardrails | 65.1 | n=93 |
| Moderation | 56.3 | n=93 |
| Code linting | 52.3 | n=90 |
| Search & retrieval | 75.7 | n=90 |
| Science | 15.7 | n=90 |
| Advertising | 41.7 | n=88 |
| Gaming | 10.5 | n=86 |
| Demand forecasting | 0.0 | n=83 |
Cells need at least 30 answered items; † fewer than 30 (indicative only; drawn as hollow, lighter points on the compare radars).
JevBench v1.6.43 values (method v1.6, before the 10 s rule)
Capability 69.95 · Composite 67.76 · rank #39 on the all-systems board · #19 open weights
intelligence 52.30 · calibration 87.61 · speed 85.87 · cost 59.35
v1.6 speed basis (124 items): p50 0.132 s · p95 0.238 s
These figures are kept for comparison. The current figures above are re-scored under the 10 s rule from the recorded per-decision times and replace them.