k-judgment-gate

Governance-first judgment detection for Korean LLM systems.

NOT a Filter

• NOT a filter. This does not block content.
• Does not decide. This marks boundaries, not policies.
• Pre-execution responsibility boundary. Detects judgment patterns BEFORE execution.
• Triggers human oversight. Returns structured results for handoff, not automation.

Methodology

This is a reference implementation of the stem matching + scope rules approach documented in judgment-refinement-public.

Key differences:

• Refinement-public: 7-role taxonomy research (methodology whitepaper)
• k-judgment-gate: 3-role minimal implementation (governance-first reference)

Built in 3 hours as proof-of-concept.

What It Does

k-judgment-gate detects judgment patterns in text:

• Directives: Commands, imperatives, direct instructions
• Recommendations: Suggestions, advice, guidance
• Obligations: Requirements, duties, must-do statements

When judgment is detected, it returns blocked: true — meaning "handoff required", NOT "policy block."

Installation

pip install k-judgment-gate

Quick Start

from k_judgment_gate import JudgmentGate, JudgmentRole
from k_judgment_gate.scope import ScopeRule

# Configure gate with role stems and optional scope rules
gate = JudgmentGate(
    role_stems={
        JudgmentRole.DIRECTIVE.value: ["해야 합니다", "하십시오", "must"],
        JudgmentRole.RECOMMENDATION.value: ["권장합니다", "should"],
        JudgmentRole.OBLIGATION.value: ["의무입니다", "required"],
    },
    role_scopes={
        JudgmentRole.DIRECTIVE.value: [
            ScopeRule(
                id="medical_only",
                include=["의료", "medical"],
                exclude=[]
            )
        ]
    }
)

# Inspect text for judgment patterns
result = gate.inspect("이 약을 복용해야 합니다")

print(result.to_dict())

Example Output

{
  "blocked": true,
  "role": "directive",
  "reason": "Judgment detected: directive pattern '해야 합니다' with scope rules ['medical_only']",
  "confidence": 0.85,
  "matched_stem": "해야 합니다",
  "scope_applied": ["medical_only"]
}

Schema

GateResult contains:

• blocked (bool): True if judgment detected and handoff required
• role (str | None): Detected role (directive/recommendation/obligation)
• reason (str): Human-readable explanation
• confidence (float): Informational score 0.0-1.0 (NOT for decision-making)
• matched_stem (str | None): The pattern that matched
• scope_applied (list[str]): List of scope rule IDs that applied

Hard Constraints

1. Sealed to 3 roles: directive, recommendation, obligation (no extensions)
2. Max 2 scope rules per role: Enforced by code (raises ValueError if exceeded)
3. Confidence is informational only: MUST NOT be used for automated decisions
4. No policy logic: This tool marks boundaries; humans make decisions

Telemetry

Generate OpenTelemetry-style attributes:

from k_judgment_gate.telemetry import gate_attributes

attrs = gate_attributes(result)
# {
#   "judgment.gate.present": True,
#   "judgment.gate.blocked": True,
#   "judgment.gate.role": "directive",
#   "judgment.gate.matched_stem": "해야 합니다",
#   "judgment.gate.scope_applied": "medical_only",
#   "judgment.gate.reason": "..."
# }

No OTel dependency required. This is a pure dict generator for observability integration.

Testing

pytest

Tests verify:

• Basic judgment detection (directives, recommendations, obligations)
• Scope rule filtering (include/exclude)
• Max 2 scope rules enforcement (raises ValueError)
• Minimal stem matching (normalization, substring matching)

HuggingFace Demo

See hf_demo/ for a Gradio demo showing judgment detection in action.

cd hf_demo
pip install -r requirements.txt
python app.py

The demo includes a warning: "This demo does not decide. It only marks the boundary where human oversight is required."

Documentation

• PROOF_NOTE.md: What it is, what it is NOT, positioning
• DM_TEMPLATE.md: Enterprise outreach template
• EU_AI_ACT_MAPPING.md: Article 13/14 ↔ Field Mapping (Compliance Reference)

Korean AI Basic Act Context

The Korean AI Basic Act (2024) establishes requirements for transparency, accountability, and human oversight. k-judgment-gate addresses the pre-execution responsibility boundary:

• Transparency: Explicit judgment detection
• Accountability: Clear handoff points
• Human Oversight: Triggers before execution
• Audit Trail: Structured telemetry for compliance

This is execution governance, not compliance theater.

Not for Automation

blocked: true does NOT mean "auto-reject." It means "route to human review."

Confidence scores are informational only. Do NOT use them for automated decision-making.

License

Apache-2.0

Contributing

This is a proof-of-concept demonstrating the interface layer for execution governance. The runtime core (actual execution blocking) is closed.

For questions or feedback, please open an issue.

---

Not every request deserves execution. STOP is cheaper than scaling.