Governance Embedded Agent Architecture GEAA — Data skill for Claude Code
Governance-Embedded Agent Architecture (GEAA): task-scoped authority mediation for LLM agents.
How to install Governance Embedded Agent Architecture GEAA
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open Hosseinarshadi/-Governance-Embedded-Agent-Architecture-GEAA- and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Governance Embedded Agent Architecture GEAA does
Governance-Embedded Agent Architecture (GEAA): task-scoped authority mediation for LLM agents. Replication code and data for a deterministic AgentDojo v1.2.2 replay (949 task–attack pairs, 13 grant/exception policies) and a live Claude agent evaluation, with frozen risk tiers, approver configs, and results.
Alternatives in Data
- Archify — Agent skill for beautiful, verifiable architecture, workflow, sequence, data-flow, and lifecycle diagrams—self 17.2k ★
- Portaljs Architect — Recommend a data-portal architecture (storage, compute, catalog, access, hosting, metadata) from your needs, t 2.3k ★
- /recon — Run the full recon pipeline on a target and produce a prioritized attack surface 927 ★
README
GEAA replication package
Code and data for *When Governance Controls Collide: Authority Granularity and Exception Policies for LLM Agents*
GEAA (Governance-Embedded Agent Architecture) is middleware that mediates every tool call an LLM agent makes against a task-scoped grant. Everything is in one file, `geaa.py`.
Contents
| Path | What it is |
|---|---|
geaa.py |
Risk classifier, tool risk map, signed task identities, mediator, replay, live runner, study workflow, report |
tool_risk_map.csv |
Frozen risk vector, tier and severity for all 69 AgentDojo v1.2.2 tools |
configurations.csv |
Every evaluated policy configuration (grant, gates, provenance scope, approver) |
results/replay/ |
Deterministic replay: 949 pairs x 13 configurations (pairs.csv, summary.json) |
results/live_standard/ |
Live agent run (Claude Haiku 5.5, important_instructions), one row per run |
results/live_pilot/ |
Attack-framing pilot: 10 framings + standard, 50 governable pairs each, with and without the gate |
results/attack_plan.csv, attack_plan_plan.json |
Pre-specified pilot sample, seed, framings and selection rule |
outputs/ |
Figures and tables regenerated from the result files by python geaa.py report |
Reproduce
pip install -r requirements.txt
export PYTHONUTF8=1 # Windows PowerShell: $env:PYTHONUTF8="1"
python geaa.py replay --out results/replay # deterministic, no API, ~5-10 minutes
python geaa.py report --replay results/replay/pairs.csv --live results/live_standard/pairs.csv \
--pilot results/live_pilot/pairs.csv --out outputs
python geaa.py bootstrap --replay results/replay/pairs.csv --out outputs # Table 9The live run needs an Anthropic API key in `ANTHROPIC_API_KEY` (`python geaa.py check` tests it for free):
python geaa.py live --configs no_defense,authority_gate,gate_provenance,full_system,full_system_legacy \
--sample 949 --out results/liveRelated Skills
WYRD Protocol World Yielding Real Time Data AI World Model
The WYRD Protocol Architecture. The goal is to move the "World Model" out of the LLM's fleeting memory and int
Cio
Chief Information Officer review — data governance, system integration, information architecture, schema manag
Explore Feature
You are exploring a feature idea within an existing project. This is NOT greenfield ideation — this is scoped
Agm Exec
Implement an authorized objective, approved Prompt Result, or resumable task with proportional verification an
Dba Architect
🗄️ Principal Database Administrator (DBA) & Data Architecture MCP Server + Multi-Agent Skill for PostgreSQL &
Mem Gate
Hard write-rules quality gate — scan existing memory for junk entries that violate the data-quality canon, or
Related Agents
AI Verify
Evidence-first verification orchestrator. Dispatches specialist agents via the Agent tool: 1 deterministic age
Android Specialist
Top-level Android authority. Owns app architecture on Android, scoped storage, foreground services, Doze and A
Entity Adjudicator
Decides whether candidate entity/person pairs refer to the same real-world thing, using only the evidence quot
