Clousight Bench

云计算指北 / Clousight 出品的云产品可复现测评框架. Clousight Bench measures the platform runtime engineering of managed cloud products — session hosting, tool-failure recovery, trace completeness, cost attribution — not model intelligence. The core orchestrates one lifecycle:

provision → setup → execute → collect → teardown → score → report

Everything product-specific is a plugin.

The reproducibility contract

Every number is classified before you trust it, and every result record carries config_hash + runner_version + evidence_layer:

Evidence layer	Meaning	Reproducibility
C	Controlled-variable measurement	Precisely reproducible
B	Environment observation	Method reproducible; numbers vary by environment
A	Documentation reading	Vendor-stated, not measured
D	Marketing material	Never load-bearing

Results are always reported per dimension — never a blended score.

Quick start (no cloud account)

pip install clousight-bench
csbench list --verbose
csbench run --domain agent-runtime --task T1.3 --platform local-sim
csbench report

Optional time-series store (Parquet + DuckDB): pip install clousight-bench[store]

Learn more

Architecture

The lifecycle, plugin contracts, and evidence model.

Extending (plugins)

Add a platform, dimension, domain, or workload engine.

Querying results

DuckDB-backed analytics over the Parquet store.

API reference

Core data schema and plugin base classes.

⌘I

Get started

Guides

Reference

The reproducibility contract

Quick start (no cloud account)

Learn more

Architecture

Extending (plugins)

Querying results

API reference

​The reproducibility contract

​Quick start (no cloud account)

​Learn more

Architecture

Extending (plugins)

Querying results

API reference

The reproducibility contract

Quick start (no cloud account)

Learn more