jg-benchmark-ops

name	jg-benchmark-ops
description	Benchmark collection and evaluation workflow for agent model assignment reviews. Use when pulling benchmarks, evaluating cost/performance, or deciding which models to use for which agents.

JG Benchmark Ops

When to Trigger

• New model release available for any agent in the project
• User requests benchmark collection, cost/performance evaluation, or model assignment review
• Periodic review (e.g. quarterly) of agent model assignments

Collection Workflow

1. Identify sources
   Use project-defined or common sources: LiveBench, SWE-Bench, Artificial Analysis, or other leaderboards. WebSearch for latest URLs and dates.

2. Fetch and parse
   WebFetch for page content. If empty or JS-rendered, use browser MCP or project fallback. Parse into a structured format (YAML/JSON). Record source URL and retrieval date for every score.

3. Store
   Write to project-defined path (e.g. benchmarks/snapshots/YYYY-MM-DD.yaml). Never overwrite; use new timestamped filename if same-day file exists.

4. Validate
   Run project's schema validator (e.g. python scripts/benchmark_schema.py --validate <path>) before considering the snapshot complete.

Evaluation Workflow

• If the project has an eval script (e.g. make benchmark-eval, scripts/benchmark_evaluate.py): run it and read the output. Use its verdicts and metrics in the report.
• If not: combine snapshot data with model pricing; for each agent, compare current model to alternatives on primary metrics; assign verdict (Excellent / Correct / Monitor / Tune / Upgrade) and note cost impact.

Verdict Definitions

Verdict	Meaning
Excellent	Current model leads its cost tier; no change needed.
Correct	Adequate; within ~5% of tier leader, no cheaper winner.
Monitor	Trails leader by ~5–15% or cheaper option within ~3%; schedule review.
Tune	Same-cost or cheaper model outperforms by >5%; recommend change.
Upgrade	Higher-cost model outperforms on critical-path role; recommend only if cost justified.

Cost and Performance

• Use input/output pricing (per token or per MTok) from provider docs or analysis sites.
• Compare: same-tier alternatives (same cost band), cheaper tier, premium tier.
• In recommendations, state: current model, suggested model, metric delta, cost delta.

Output Expectations

• Collection: Sources, dates, snapshot path, list of models collected (and any missing).
• Evaluation: Table (Agent | Model | Verdict | Key metrics); Recommendations (agent, change, before/after, cost impact).
• No agent or rule file updates unless the user explicitly asks to apply recommendations.

Anti-Patterns

• Do not record scores without source URL and date.
• Do not overwrite existing snapshots.
• Do not apply model assignment changes without explicit approval.
• Do not skip schema validation when the project defines one.