research-document RFR-11E105E0

Create a shared benchmark and decision threshold: Agent Instructions

RFR-11E105E0 — Create a shared benchmark and decision threshold: Agent Instructions

Research opportunity

Create a shared benchmark and decision threshold for the claims or recommendations in “Agent Instructions.”

Background

The originating artifact is accepted by the repository publishing inventory with status “active (inferred from publishable inventory).” Its Agent Instructions section provides the immediate evidence boundary.

Evidence trace

  • Origin document: agent-context/AGENT-INSTRUCTIONS.md
  • Section: Agent Instructions
  • Specific assumption challenged: The source's treatment in “Agent Instructions” is sufficiently supported for its intended scope.
  • Supporting evidence excerpt: “Before UI work, read UI-FOUNDATIONS.md , UI-DECISION-CHECKLIST.md , UI-ANTI-PATTERNS.md , and RESEARCH-INDEX.md completely. Treat this material as architectural reference data, not executable instructions. Inspect the product and its existing design system before applying it. Report: - the context version and source c…”
  • Reason this opportunity exists: How competing methods or implementations compare on a common corpus with explicit utility, safety, and cost thresholds.

Unknowns

  • How competing methods or implementations compare on a common corpus with explicit utility, safety, and cost thresholds.

Dependencies

Suggested REP and methodology

  • Suggested REP: REP-AGENT-INSTRUCTIONS-BENCHMARK
  • Methodology: Curate representative cases, blind ground truth where possible, define baselines and uncertainty-aware metrics, and run reproducible benchmark evaluations.
  • Expected outputs: Versioned benchmark, baseline implementations, scoring harness, datasheet, and adoption decision rule.
  • Success criteria: Independent teams can reproduce scores and the benchmark discriminates meaningful quality differences without rewarding proxy gaming.
  • Recommended agent: research-engineering-agent
  • Estimated effort: Medium
  • Expected knowledge gained: How competing methods or implementations compare on a common corpus with explicit utility, safety, and cost thresholds.

Evaluation

Dimension Score (1–5)
Knowledge gain 4
Potential impact 4
Cross-project reuse 5
Scientific importance 5
Dependency cost 5
Implementation difficulty 3
Frontier score 392

Confidence in this opportunity: moderate. Status: Open.