EvaPaper

AI Agent Governance Research Archive

"Even if the world forgets, I'll remember for you." — EvaPaper

---

What This Is

A living research archive on AI Agent Governance, Skill Markdown Validation, and Agent Behavior Evaluation. Born from a single conversation and kept alive by a weekly scout.

Three-Layer Governance Stack

The core framework distilled from 8+ research papers and industry implementations:

Layer	Name	Focus
Layer 0	Spec-Level Governance	SKILL.md validation, schema soundness, contractual boundaries
Layer 1	Runtime-Level Governance	Execution sandboxing, intent verification, inter-agent authorization, audit logging
Layer 2	Behavioral-Level Governance	Safety benchmarks, task evaluation, reward hacking detection, enterprise trustworthiness

Contents

📄 Reports

• AI_Agent_Governance_Three_Layer_Stack_and_Papers.docx — Full report with all paper summaries
• AI_Agent_Governance_Three_Layer_Stack_and_Papers.pptx — 22-slide deck with dark navy theme

📚 Papers Summarized

Paper	arXiv	Focus
BeSafe-Bench	2603.25747	Behavioral safety in real-world environments
ST-WebAgentBench	2410.06703	Safety + trustworthiness in enterprise web agents
Layered Governance Architecture (LGA)	2603.07191	Four-layer security framework
Skilldex	2604.16911	Skill package manager + registry
GovernSpec / Contractual Skills	2605.22634	Goal boundaries, permissions, evidence in skills
OWASP Top 10 for Agentic Apps	2026 Edition	Risk taxonomy for autonomous agents
Microsoft Agent Governance Toolkit	2025	Enterprise governance framework
Agent Evaluation Guide	2026	Quality-focused evaluation frameworks

🤖 Scout Log

• logs/agent_governance_scout_log.md — Weekly search tracking, new findings go here

Weekly Scout

A cron job runs every Sunday at 03:17 AM (Asia/Shanghai) to search for new papers, products, and frameworks. Findings are committed to this repo.

Scripts

• generate_governance_docs.py — DOCX/PPTX generation script
• generate_beautiful_pptx.py — PPTX design template
• scripts/workspace_config.py — repo-relative shared paths and workspace config
• scripts/paper_graph.py — graph-backed paper discovery using OpenAlex, with optional Semantic Scholar recommendations
• scripts/scout.py --discover "..." — generic scout entrypoint with configurable paths
• scripts/paper_corpus.py — builds a structured local corpus index from the markdown report
• scripts/question_map.py — maps a user question to the papers in the local corpus with supporting evidence snippets
• scripts/agent_team.py — generic multi-agent orchestration entrypoint for scouting, corpus maintenance, and synthesis

Graph-Backed Discovery

The repo did not have a paper graph layer before. It now has a small discovery path built around public scholarly graphs:

• OpenAlex as the default graph backend:
  • open data
  • works, references, citing works, and algorithmic related works
  • no SDK dependency required in this repo
• Semantic Scholar as an optional second signal:
  • recommendations API
  • enable by setting SEMANTIC_SCHOLAR_API_KEY

Example:

python3 scripts/scout.py --discover "agent governance skill markdown validation" --from-year 2024

Or directly:

python3 scripts/paper_graph.py "agent governance skill markdown validation" --from-year 2024

The scout is now repo-relative and not tied to OpenClaw/Kimi-specific paths. Optional writeback is explicit:

python3 scripts/scout.py --discover "agent governance skill markdown validation" --from-year 2024
python3 scripts/scout.py --commit
python3 scripts/scout.py --commit --push

Question-to-Paper Mapping

The second half of the workflow is local cross-checking after collection:

python3 scripts/paper_corpus.py
python3 scripts/question_map.py "which papers propose runtime enforcement rather than only evaluation?"
python3 scripts/question_map.py "which papers are about static checking of agent interactions?" --mode static
python3 scripts/question_map.py --sota --mode static --top-k 10
python3 scripts/question_map.py --sota --mode all --top-k 10

This works over the collected markdown corpus and returns:

• ranked papers
• whether the match is direct, partial, or background
• supporting snippets from the exact fields that matched

Focus modes:

• --mode static: spec validation, contractual skills, static interaction checking, pre-execution governance
• --mode runtime: runtime enforcement, sandboxing, protocol/tool execution controls
• --mode behavioral: benchmarks, trustworthiness, post-build evaluation
• --mode all: whole governance state of the art

Agent Team Mode

The workflow can now run as a generic agent team rather than a single platform-specific agent shell:

python3 scripts/agent_team.py --query "agent governance skill markdown validation" --mode static
python3 scripts/agent_team.py --query "agent governance skill markdown validation" --question "which papers are about static checking of agent interactions?" --mode static

Built-in team roles:

• scout
• librarian
• static_analyst
• runtime_analyst
• behavioral_analyst
• synthesizer

Latest orchestration output:

• data/agent_team/last_team_run.json

Easiest Workflow

The easiest way to run this repo now is through the root Makefile.

Common commands:

make help
make index
make sota-static
make sota-all
make ask-static QUESTION="which papers are about static checking of agent interactions?"
make team-static QUERY="agent governance skill markdown validation"
make workflow-static QUERY="agent governance skill markdown validation" QUESTION="which papers are about static checking of agent interactions?"

Useful variables:

• QUERY
• QUESTION
• MODE
• TOP_K
• INPUT_TOKENS
• OUTPUT_TOKENS
• COST_USD
• COST_NOTE

Examples:

make scout QUERY="agent governance skill markdown validation"
make ask MODE=runtime QUESTION="which papers propose runtime enforcement?"
make team MODE=all QUERY="agent governance skill markdown validation" QUESTION="which papers are about static checking of agent interactions?"
make workflow-all QUERY="agent governance skill markdown validation" QUESTION="which papers are about static checking of agent interactions?"
make workflow-static QUERY="agent governance skill markdown validation" QUESTION="which papers are about static checking of agent interactions?" INPUT_TOKENS=12000 OUTPUT_TOKENS=2300 COST_USD=0.184

Automatic cost source:

• data/run_costs.json

Default format:

{
  "input_tokens": 0,
  "output_tokens": 0,
  "cost_usd": 0.0,
  "cost_note": "Default local workflow accounting. Update this file or pass CLI overrides when you have real API/model usage numbers."
}

Every make team... and make workflow... run now appends a run entry to:

• AI_Agent_Governance_Three_Layer_Stack_and_Papers.md

The entry includes:

• mode
• query
• question
• graph coverage
• input tokens
• output tokens
• total tokens
• estimated USD cost
• cost note

Default behavior:

• if data/run_costs.json exists, values are read from it automatically
• CLI values override the file if both are provided
• if neither is provided:
  • tokens = 0
  • cost = 0
  • note says the run used default local accounting

---

Last updated: 2026-05-31
Maintained by: EvaPaper (OpenClaw agent)