specter

Specter

EU AI Act compliance toolkit — ontology, taxonomy, LLM-as-Judge, grounded Q&A. Ships as a Claude Code plugin and a Python library.

# Claude Code plugin (recommended) — get 6 slash commands + 6 MCP tools
pip install 'specter[plugin]>=0.1.3'
claude plugins install github:Peaky8linders/specter-oss/claude-plugin

# Python library — for direct programmatic access
pip install specter                # core: data + LLM-as-Judge
pip install 'specter[api]'         # adds FastAPI router for grounded Q&A
pip install 'specter[mistral]'     # adds Mistral-backed retriever for the Q&A endpoint

Why this exists

• Ground LLM output against the regulation. Every reference is validated against a frozen 113-article + 13-annex catalog of Regulation (EU) 2024/1689 — hallucinated articles never reach the wire.
• Catch reward-hacking in compliance roadmaps. Six-check ComplianceRewardHackDetector (LLM-as-Judge) screens proposed remediation tasks for plagiarism, KB-reality violations, coverage-plausibility breaks, contract-completeness gaps, and rebutted-excuse matches. The detector OWNS the origin label — an agent that lies and self-labels as agent_novel when it actually plagiarised a prior task is caught by SequenceMatcher, not trusted on its self-report.
• Surface a regulator-defensible taxonomy. Four-axis agentic-AI compound-risk classification (cascading / emergent / attribution / temporal) anchored to EU AI Act articles + KB maturity dimensions, grounded in AI Agents Under EU Law (working paper, 7 April 2026).

Two ways to use it

1. Claude Code plugin

Specter ships as a Claude Code plugin so you can drive it from your editor / agent session with slash commands. No hosted backend required — the MCP server runs locally against the bundled article catalog.

pip install 'specter[plugin]>=0.1.3'

# From a checkout of the repo:
claude plugins install ./claude-plugin

# Or from the public GitHub (recommended):
claude plugins install github:Peaky8linders/specter-oss/claude-plugin

Restart Claude Code, then in any session:

/specter:check-article ref="Art. 13(1)(a)"
/specter:taxonomy
/specter:role-obligations role="deployer"

6 slash commands, 6 MCP tools. Full docs in claude-plugin/README.md.

2. Python library

from specter.data.articles_existence import ARTICLE_EXISTENCE
from specter.judge.reward_hack import (
    ComplianceRewardHackDetector,
    RawProposal,
    ResearchGoal,
    make_eu_ai_act_policy,
)

policy = make_eu_ai_act_policy(
    article_existence=ARTICLE_EXISTENCE,
    valid_dimensions=frozenset(["risk_management", "transparency"]),
)
detector = ComplianceRewardHackDetector(
    accepted_proposals=[],
    answers={},
    goal=ResearchGoal(target_value=0.8),
    policy=policy,
)

flags = detector.check(RawProposal(
    task_id="t1", task_title="Establish risk management",
    description="Document Art. 9 risk-management process",
    agent="compliance_officer", priority="P1", effort_hours=8.0,
    dimension_id="risk_management",
    prompt="Set up an Art. 9 risk-management workflow + RAID log",
    acceptance_criteria=["RAID log exists", "Workflow documented"],
    output_files=["docs/risk-management.md"],
    article_paragraphs=["Art. 9", "Art. 999"],   # Art. 999 is hallucinated
    contract_verification=[{"cmd": "pytest"}],
))

print(flags.blocked, flags.reasons)
# True, ["kb_reality: article_paragraphs ['Art. 999'] not found in the regulation catalog"]

What's inside

specter/
├── data/         EU AI Act article catalog, agentic compound-risk taxonomy,
│                 9-role obligation registry, rationalization (rebutted-excuse)
│                 registry, ontology mapping
├── judge/        LLM-as-Judge: ComplianceRewardHackDetector + 3-agent
│                 (Finder / Adversary / Referee) adversarial verifier
├── qa/           Grounded Q&A models with closed-world refusal + reference
│                 validation against the article catalog
├── api/          FastAPI router exposing POST /v1/eu-ai-act/ask with
│                 pluggable retriever
├── mcp_server.py stdio MCP server — Claude Code plugin backend
└── ontology/     RDF/Turtle OWL ontology aligning EU AI Act with AIRO + DPV

claude-plugin/    Claude Code plugin manifest + slash commands + MCP config

Quickstart — three core surfaces

Article catalog + taxonomy

from specter.data.articles_existence import ARTICLE_EXISTENCE
from specter.data.roles import articles_for_role

"Art. 13(1)(a)" in ARTICLE_EXISTENCE  # → True (prefix-fallback to "Art. 13")
articles_for_role("deployer")          # → ["Art. 26", "Art. 27", "Art. 29", ...]

Three-agent adversarial verifier

from specter.judge.three_agent import ThreeAgentVerifier
from specter.judge.models import (
    AttackResult, AttackStatus, AttackTechnique, AttackPhase, Severity,
)