Try Hanzo

Benchmarks / Kai / AG News

Kai · Decision harness

AG News

Kai and Jev on the same 400 questions of AG News, one run per Kai revision.

Latest

History

12 results across 2 metrics, newest first. Superseded runs stay, so a revision's progress can be read.

85%91%96%09-2709-2810-0310-16Kai a5 · df1c16a2: 95.0%, Sep 27, 2026Kai a8 · 0cfeef05: 94.3%, Sep 28, 2026Kai a7 · 0834a74f: 94.3%, Sep 28, 2026Kai a6 · a211cc70: 94.0%, Sep 28, 2026kai-1 · 0834a74f: 94.3%, Oct 3, 2026kai-1.4 · c1856a77: 93.8%, Oct 16, 2026
MeasuredSubjectMetricValueBaselineResultStatus
Oct 16, 2026kai-1.4 · c1856a77Accuracy93.8%Jev 86.0%
Win
Live · Hanzo-measured
Oct 16, 2026kai-1.4 · c1856a77Calibration error (ECE)0.022Jev 0.102
Win
Live · Hanzo-measured
Oct 3, 2026kai-1 · 0834a74fAccuracy94.3%Jev 86.0%
Win
Superseded · Hanzo-measured
Oct 3, 2026kai-1 · 0834a74fCalibration error (ECE)0.033Jev 0.102
Win
Superseded · Hanzo-measured
Sep 28, 2026Kai a8 · 0cfeef05Accuracy94.3%Jev 86.0%
Win
Superseded · Hanzo-measured
Sep 28, 2026Kai a7 · 0834a74fAccuracy94.3%Jev 86.0%
Win
Superseded · Hanzo-measured
Sep 28, 2026Kai a6 · a211cc70Accuracy94.0%Jev 86.0%
Win
Superseded · Hanzo-measured
Sep 28, 2026Kai a8 · 0cfeef05Calibration error (ECE)0.024Jev 0.102
Win
Superseded · Hanzo-measured
Sep 28, 2026Kai a7 · 0834a74fCalibration error (ECE)0.034Jev 0.102
Win
Superseded · Hanzo-measured
Sep 28, 2026Kai a6 · a211cc70Calibration error (ECE)0.028Jev 0.102
Win
Superseded · Hanzo-measured
Sep 27, 2026Kai a5 · df1c16a2Accuracy95.0%Jev 86.0%
Win
Superseded · Hanzo-measured
Sep 27, 2026Kai a5 · df1c16a2Calibration error (ECE)0.070Jev 0.102
Win
Superseded · Hanzo-measured

Conditions

split: harness400 questions, 400 answereddataset dde140a63fc56 Kai revisions measuredJev: typesafe/jev-1.13-20260917, measured Sep 25, 2026

Hardware

CPU AS SERVED

Source

research run jev.ag_news-harness-kai-1.4.c1856a77-kai

Reproduce

bench score --harness . --kai results/kai/preds.json.gz --out results/scores.json # in hanzoai/benchmarks/decision @ 5f616dd

More Kai benchmarks

DAIR Emotion · Banking77 · Support triage · Email spam · Phishing · Toxicity · RAG relevance · Model routing · Typed decisions · MASSIVE · Choosing among many options · Latency of one decision

Every result for AG News on the shelf · How these are measured

Build what’s next.