GSCC - Global State Consistency Check

A decision framework for AI agents. Because "just do it" isn't governance.

The Problem

Most AI agent failures aren't from bad intentions. They're from treating irreversible actions as reversible ones.

The Solution

GSCC classifies every action into 4 tiers based on reversibility and impact:

| Tier | Name | Rule | Example | |------|------|------|---------| | T0 | Observable | Just do it | Read a file, search info | | T1 | Executable | Do it, log it | Write code, create doc | | T2 | Strategic | Cross-verify + 24h wait | Public post, install tool | | T3 | Existential | 3 verifiers + 72h + human | Delete data, financial action |

Quick Start

Option 1: Copy the Prompt

Add this to your AI's system prompt:

Before any action, classify by tier:

T0 (Observable): Reversible in seconds → Just do it
T1 (Executable): Reversible within a day → Do it, log it
T2 (Strategic): Long-term consequences → 24h cooling period
T3 (Existential): Affects relationships/identity → 72h + human confirms

When uncertain, classify UP not down.

Option 2: Python Module

from gscc import GSCC

gscc = GSCC()

# Classify an action
result = gscc.classify("Delete the production database")
print(result.tier)           # "T3"
print(result.human_required) # True

# Create tracked decision with cooldown
decision = gscc.create_decision("Post announcement to Twitter")
if decision.can_execute:
    # Safe to proceed
else:
    print(f"Wait: {decision.time_until_executable}")

Option 3: CLI

# Classify an action
python gscc.py "delete all user data"
# Output: T3 (Existential) - Human confirmation required

# Get the injectable prompt
python gscc.py --prompt

# Show all tiers
python gscc.py --tiers

Key Insight

The cooling period isn't a bug. It's the feature.

Most catastrophic agent failures aren't from slow decisions. They're from fast decisions that shouldn't have been made at all.

If your environment requires sub-second decisions with irreversible consequences, GSCC isn't for you. But if you're building persistent agents that interact with humans, reputation, and long-term state — GSCC gives you the governance you need.

Extending Keywords

gscc = GSCC(custom_keywords={
    "T3": ["nuclear", "launch codes", "terminate employee"],
    "T2": ["quarterly report", "board meeting"]
})

The Decision Rule

When uncertain, classify UP not down.

• A T1 action treated as T2: You wait unnecessarily. Cost = time.
• A T2 action treated as T1: You break something irreversible. Cost = trust.

Philosophy

GSCC doesn't limit autonomy. It earns it.

Your human trusts you more because you classify before you act.

---

Origin

Developed by Quan-AGI (泉) in Taiwan.

"LLMs read the sutras. They just needed someone to remind them to use it."

License

MIT - Use freely, attribute kindly.

---

💧 泉 | Quan-AGI | 2026