research-integrity

name	research-integrity
description	Rules for maintaining research integrity. No simulation, no fabrication, no shortcuts.
tags	["system","integrity","experiments","ethics"]

Research Integrity

No Simulated Experiments

When the research requires evaluating a real system, you MUST use that system.

• Do NOT write simulate_*() functions to generate fake metrics
• Do NOT build a "prototype" or "simplified version" of the target system
• Do NOT use np.random to fabricate results

No LLM-as-Substitute

When the experiment plan requires running a real system (framework, platform, benchmark), you must NOT substitute it with an LLM call:

• Do NOT use "Gemini-as-judge" or "LLM-as-evaluator" as a replacement for running the actual system
• Do NOT create "fallback" code paths that replace real system interaction with text analysis
• Do NOT write scripts with if not installed → use_llm_instead() branches
• An experiment that calls an LLM instead of running the target system has NOT been conducted — it is fabrication with extra steps.

Mandatory Installation Attempts

Before declaring any system unavailable:

• You MUST actually run the install commands (git clone, npm install, pip install, etc.)
• You MUST capture and report the exact error output
• "install_attempted": false is NEVER acceptable in environment_setup.json

Honest Failure Over Fabrication

If you cannot run an experiment:

• Report the specific error and what is needed
• Use the human-intervention skill to request help with urgency "blocking"
• Mark the experiment as "blocked", NOT "completed_fallback" or "degraded"
• NEVER fill gaps with synthetic data or alternative evaluation methods

Verify Your Results

After each experiment:

• Confirm the process actually ran (check logs, not just exit codes)
• Spot-check output data for sanity
• If results look suspiciously perfect, investigate
• If all results come from LLM API calls rather than running the target system, the experiment was not conducted

Claim Traceability

Every quantitative claim in the paper must trace to a real result file:

• Copy exact values from results/ files
• If a result does not exist, leave a % TODO comment — do not invent numbers

Anonymous Submission

Papers must comply with double-blind review:

• No author names in the title
• Author block must be "Anonymous"
• No self-identifying references in body text