Argus Validation Protocol v1.0

Argus is an execution observatory protocol for reproducible validation.

It does not claim absolute performance. It records whether structural behavior changes are repeatedly observable under identical conditions.

This repository is the public documentation and protocol record for Argus. Public binaries are distributed through GitHub Releases. The internal implementation source is not currently published in this repository.

For execution-level variability and redundant-computation studies, Argus v1 provides execution_instability mode. This mode keeps the standard v1 record and adds instability_metrics.json for path, reuse, recomputation, and work-inflation observations.

What Argus Outputs

Argus output is a Reproducible Observation Record:

• metrics.json
• report.md
• run_meta.json
• instability_metrics.json (when mode: execution_instability)

Optional submission package:

• argus_result.zip
• argus_result.sanitized.zip (--sanitize)

Try it in 10 seconds

Linux (amd64):

curl -L https://github.com/tongro2025/Argus/releases/latest/download/argus-linux-amd64 -o argus
chmod +x argus
./argus --help

macOS (Apple Silicon):

curl -L https://github.com/tongro2025/Argus/releases/latest/download/argus-macos-arm64 -o argus
chmod +x argus
./argus --help

Windows binaries are also published as release assets.

Protocol CLI Contract (v1)

argus doctor
argus run <config.yaml>
argus run <config.yaml> --mode execution_instability
argus report <run_dir>
argus export <run_dir>
argus export <run_dir> --sanitize

The commands above define the public protocol interface. Current public release binaries include execution_instability mode support.

Example config.yaml

seed: 42
steps: 500
repeat: 5
warmup_steps: 20
mode: standard
ricci: [off, on]
workload:
  nodes: 8
  requests_per_step: 128

Warmup steps are executed but excluded from final statistics.

With mode: execution_instability, Argus additionally writes instability_metrics.json next to the standard artifacts.

What gets measured

• P95 latency, P99 latency, throughput
• migration / retry / invalidation count
• error_rate, oom_count, deadlock_count, crash_count
• peak_memory_usage, avg_memory_usage
• per-metric Mean, Std Dev, Min, Max

With mode: execution_instability, Argus also records:

• distinct execution path count and ordering changes
• reuse opportunities, reuse failures, and intermediate rebuild count
• observed work, reference work, repeated operations, and work inflation ratio
• recompute ratio and execution path variability score

Environment reliability checks

argus run records runtime noise context and warns when needed:

• cpu_load_percent
• memory_usage_percent
• gpu_temperature
• gpu_utilization
• thermal_throttle_detected

If noise is high, Argus shows:

[Argus Warning]
Environment noise detected.
Results may be unreliable.
Continue? (y/n)

Sanitized export

Use argus export <run_dir> --sanitize when sharing results.

Sanitization removes/redacts:

• username
• absolute user-home path
• hostname
• env assignments
• IP address

Documentation

English:

• Architecture Update (v0.2 to v1.0)
• Execution Instability Mode
• Validation Protocol
• Experiment Rules
• Claims and Limits
• Reproducibility Guide

한국어:

• 아키텍처 업데이트 (v0.2 to v1.0)
• Execution Instability Mode
• 검증 프로토콜
• 실험 규칙
• 주장과 한계
• 재현성 가이드

License

Licensed under Apache License 2.0.

• LICENSE
• NOTICE