Live verification: September 20, 2026

Fetched 180 recent warning/error log entries from Loki. Reduced them to reviewed diagnostic templates before sending anything to Jev: generated source labels, no real hosts, addresses, paths, image digests, headers, or credentials. Unrelated entries were excluded.

The retained snapshot had 156 entries across seven sources. It was served through a local Loki-format HTTP endpoint so all three runs saw identical input. The CLI used the real TypeSafe SDK and API; jev-latest resolved to jev-1.13.0.

Run	API calls	Input tokens	Elapsed
Empty cache	7	8,789	1.005 s
Warm cache	0	0	0.010 s
Cache disabled	7	8,789	1.009 s

All runs exited successfully with no batch errors. Warm-cache decisions matched the cold run exactly. Uncached decision labels also matched:

• Endpoint deprecation: suppress.
• RSS 304 responses: watch.
• Registry lookup, image-signature, release-schema, scrape-configuration, and notification-delivery failures: notify.

These are observed model judgments, not verified accuracy labels. Each time is one in-process CLI invocation, excluding initial log capture and Python startup. This is an operation check, not a general performance benchmark. Nothing was remediated or notified externally.

Public sample

logtriage/fixtures/observed.json contains 16 representative lines, with duplicates capped at two. Diagnostic status codes and timeout values are retained; identifying context is not. The private raw capture was deleted. The reduced fixture is not enough to reproduce the exact timing or count-dependent judgments above.

The offline batching tests load both this fixture and the original demo. They do not call Jev or assert model predictions. To inspect the additional sample locally:

from logtriage.batch import build_batches, demo_fixture_path, load_demo_streams

streams = load_demo_streams(demo_fixture_path().with_name("observed.json"))
batches = build_batches(streams)

For repeatable cold/warm/uncached timing with the original bundled demo, use scripts/bench-cache.sh.