Jev Logs triage explorer

Measured routing of Jev Logs (jevlogs@0.3.0 local build) on sanitized Loghub-derived HDFS and BGL samples.

This Space never calls a model and has no API keys. Sliding retainBelow recomputes route locally: retain only when priority is low, value ≤ 25, and actionableProbability < threshold. Protected, unavailable, and missing-probability records stay analyze.

Dataset: reachjalil/jevlogs-log-triage-benchmark
GitHub: reachjalil/jevlogs
npm: jevlogs
Site: jevlogs.com

Package under test: jevlogs@0.3.0 (git a925673). Seed 20260916. Estimated Jev spend from logged tokens: $0.15393621600000001 (3665148 input tokens, 6840 calls with usage). Luna structured-output job on the same Gateway key: $0.081224. Confirm on the Vercel AI Gateway dashboard. HDFS labels are block-level, not line-level. BGL labels are line-level alerts. The sample oversamples anomalies (~30%); it is not a production mix. Cache hit rate depends on how repetitive the workload is.

E1 baseline (`retainBelow = 0.1`, cache off)

Headline metrics

Headline metrics

hdfs	2500	750	0.9933333333333333	0.0084	0.7619047619047619	0	0	987.5707500000717	1371.3091250000289	532.4099462365591


hdfs	2500	750	0.9933333333333333	0.0084	0.7619047619047619	0	0	987.5707500000717	1371.3091250000289	540.1232
bgl	2500	750	1	0.0012	1	1	1	897.747083999915	1196.0157500000205	532.4099462365591

GPT-5.6 Luna vs Jev (same 400-line slice, same Gateway key)

Structured Luna (generateObject) vs Jev on the same sanitized lines. BGL Luna unavailable rows are Gateway 503/abort, counted as analyze; all were label=normal. HDFS Luna retain is higher because it scores value ≤ 25; Jev sits just above 25.

Luna vs Jev

Luna vs Jev

hdfs	400	0.8333333333333334	0.9916666666666667	0.14	0.0075	0.9725	31


hdfs	400	0.8333333333333334	0.9916666666666667	0.14	0.01	0.865	0
bgl	400	1	1	0.02	0.0075	0.9725	31

E2 saved threshold sweep

Recall vs routing rate from saved probabilities

Recall vs routing rate from saved probabilities

hdfs	0.05	2500	0.9933333333333333	0.0084	0.7619047619047619	0.9916


hdfs	0.05	2500	1	0	null	1
hdfs	0.1	2500	0.9933333333333333	0.0084	0.7619047619047619	0.9916
hdfs	0.2	2500	0.9186666666666666	0.106	0.769811320754717	0.894
hdfs	0.3	2500	0.9186666666666666	0.106	0.769811320754717	0.894
hdfs	0.5	2500	0.9186666666666666	0.106	0.769811320754717	0.894
bgl	0.05	2500	1	0	null	1
bgl	0.1	2500	1	0.0012	1	0.9988
bgl	0.2	2500	1	0.0068	1	0.9932
bgl	0.3	2500	1	0.0116	1	0.9884
bgl	0.5	2500	1	0.0196	1	0.9804

E7 cache and E8 retain rules

E7 uses the default 1,000-entry / 5-minute cache. E8 adds three retain rules (HDFS PacketResponder terminating; BGL icache parity corrected; BGL rbs signal handler). An HDFS rule that matches a heartbeat template will also match that template on block-labeled anomalies; that is a finding about the rule, not about Jev.

Cache hit rate, model calls, rule hits

Cache hit rate, model calls, rule hits

E7 cache	hdfs	0.9648	1097	375	75	0.9933333333333333	0.0084


E7 cache	hdfs	0.9648	88	0	0	0.9933333333333333	0.0084
E7 cache	bgl	0.1568	1097	0	0	1	0.0012
E8 rules	hdfs	0.816	85	375	75	0.8933333333333333	0.1584
E8 rules	bgl	0.12	1095	94	0	1	0.0388

Recompute routing from saved probabilities

Dataset

hdfs bgl

retainBelow

0 0.5

Metrics at this threshold

Sweep including this threshold

Sweep including this threshold

Missed anomalies at this threshold (first 25)

Missed anomalies at this threshold (first 25)

Savings calculator (estimates)

Mirrors estimateSavings() in jevlogs. retainedFraction here is the share still sent to a downstream LLM.

Logs per period

Downstream input tokens / log

Downstream output tokens / log

Downstream $/M input (default: GPT-5.6 Luna)

Downstream $/M output (default: GPT-5.6 Luna)

Analyze fraction (retainedFraction)

0 1

Jev $/M input

Jev question tokens / log (measured default)

Also show nothing-filtered case

Estimated USD

Estimated USD

Spend recorded in this benchmark

Estimated from logged Jev input tokens: $0.15393621600000001. GPT-5.6 Luna structured-output job (same Gateway key): $0.081224. Jalil should confirm both against the Gateway dashboard. Jev output is priced at $0 on the model page.

Upstream logs: Loghub HDFS_v1 and BGL, research/academic license with required citation.