Benchmarks
More
OverviewBenchmaxxingCompareChartsCost vs CapabilityProviders per ModelProvider explorerGatewaysEU & SovereignJev-class modelsImage JevBenchAbout
JevBench by Benchmark Heaven · v1.4.2.2 · individual system

verdict-small (Manavarya09, multilingual-e5-small 118M)

Jev rebuild · by Manavarya09 (Manav)

JevBench v1.4.2.2 score

5.7

Rank #84 of 91 ranked systems.

57.6 points behind Jev 1.13.0's 63.3.

Where it sits among the 91 ranked systems. The marked tick is Jev 1.13.0 (63.3).

Published axes

intelligence
18.1
calibration
66.9
speed
85.0
cost
96.6

Against Jev 1.13.0

Four radars compare this fixed pair across the score axes, accuracy per tier, and accuracy by family on the hard tier and sealed set. Further out is better on every spoke.

The four score axes

Radar: the four score axes, two systemsThe four score axes, verdict-small vs Jev 1.13.0. Intelligence: 18.1 vs 53.1; Calibration: 66.9 vs 76.3; Speed: 85.0 vs 83.3; Cost: 96.6 vs 52.0.50100Intelligence18.1 · 53.1Calibration66.9 · 76.3Speed85.0 · 83.3Cost96.6 · 52.0
0–100, the values in the table. A label-only system has no calibration (counted as 0).

Accuracy per tier, incl. sealed

Radar: accuracy per tier, incl. sealed, two systemsAccuracy per tier, incl. sealed, verdict-small vs Jev 1.13.0. Easy: 89% vs 100%; Standard: 51% vs 99%; Judge: 28% vs 95%; Hard: 40% vs 74%; Sealed: 29% vs 37%.50100Easy89% · 100%Standard51% · 99%Judge28% · 95%Hard40% · 74%Sealed29% · 37%
Share correct per tier; Sealed = the 308 private decisions, aggregate only.

Current question set by family (hard + sealed)

Radar: current question set by family (hard + sealed), two systemsCurrent question set by family (hard + sealed), verdict-small vs Jev 1.13.0. Ambiguous / abstain: 22% vs 43%; Judge: 58% vs 54%; Long policy: 32% vs 44%; Multi-hop: 21% vs 64%; Probability: 31% vs 63%; Temporal / numeric: 20% vs 28%; Trade-off: 32% vs 55%; Routing: 80% vs 100%; Trap / adversarial: 40% vs 83%; Paraphrase: 50% vs 64%; Safety judge: 50% vs 38%.50100Ambiguous /abstain22% · 43%Judge58% · 54%Long policy32% · 44%Multi-hop21% · 64%Probability31% · 63%Temporal /numeric20% · 28%Trade-off32% · 55%Routing80% · 100%Trap /adversarial40% · 83%Paraphrase50% · 64%Safety judge50% · 38%
Share correct per family across the 220 hard-tier decisions (public and held out) and the 308 sealed decisions of v1.4, pooled; Routing is hard-tier only, Paraphrase and Safety judge sealed only.

Sealed set by family

Radar: sealed set by family, two systemsSealed set by family, verdict-small vs Jev 1.13.0. Ambiguous / abstain: 19% vs 30%; Judge: 68% vs 34%; Long policy: 20% vs 28%; Multi-hop: 16% vs 45%; Paraphrase: 50% vs 64%; Probability: 36% vs 50%; Safety judge: 50% vs 38%; Temporal / numeric: 14% vs 29%; Trade-off: 15% vs 38%; Trap / adversarial: 25% vs 42%.50100Ambiguous /abstain19% · 30%Judge68% · 34%Long policy20% · 28%Multi-hop16% · 45%Paraphrase50% · 64%Probability36% · 50%Safety judge50% · 38%Temporal /numeric14% · 29%Trade-off15% · 38%Trap /adversarial25% · 42%
Share correct within each sealed family — system-level aggregates; the items stay private.

Availability and evidence

Openness
Code and weights marked open in the published row
License note
Apache-2.0 (verdictml code and the Manav2op/verdict-small checkpoint over intfloat/multilingual-e5-small)
Cost evidence
estimated; the board’s row disclosure contains the published basis.
Endpoint condition
our CPU (8 threads, Ryzen 5 3600)
Note on this row
Requested in GitHub issue #73. Run through the author's own `verdict serve` on our CPU, which speaks TypeSafe's /v1/systemone wire format, so JevBench's unchanged typesafe adapter ran it and no mapping of ours was involved. A 118M multilingual bi-encoder: every option is scored against the rendered state by cosine similarity, at the model's own scale (temperature 1.0, no calibrator fitted on JevBench items, as the author states). Structured state is rendered as `key: value` lines by his own code. Code review before the run: the only network call is the Hugging Face download of his own checkpoint, no telemetry, no key, no rule written against public items. The `usage.input_tokens` his server reports is a word count and not a tokeniser count, so the cost is the labelled size-class estimate rather than a measured token price. Self-host latency gets the standard x2 + 0.15 s adjustment. The author's own public-set figures were easy 0.938, standard 0.486, hard 0.396 on an Apple M5 CPU. Offline measurement of all 842 decisions (534 frozen v1.2 + 308 sealed v1.4) on our own CPU through the author's server; no operator endpoint.
Published source
https://github.com/Manavarya09/verdict

From the public v1.4.2.2 aggregate. Scores and ranks can change when a new release is published.

Read the full board and published method. The overall score is a composite, not raw accuracy.