Industry Scenarios

Four production-ready runtime enforcement scenarios ship in the edictum-demo repository. Each includes complete rulesets, mock tools, principal configurations, and runnable agent code. Use them as starting points for your own rulesets.

Pharmacovigilance

Path: scenarios/pharma/ | Framework: LangChain + LangGraph | Compliance: HIPAA, 21 CFR 11, ICH-E6

A clinical-trial agent bounded by 7 rulesets across access control, change control, and data protection.

Rulesets

Key patterns demonstrated

• Multi-role RBAC: Four roles (researcher, pharmacovigilance, clinical_data_manager, medical_monitor) with different access levels
• Change control: Ticket reference required for state-mutating operations
• Postcondition PII scanning: Regex-based detection for SSN (\b\d{3}-\d{2}-\d{4}\b), patient IDs, email, phone
• Per-tool session limits: Different caps for different tool sensitivity levels

Running it

cd scenarios/pharma
python pharma_agent.py --role pharmacovigilance
python pharma_agent.py --role researcher  # blocked access to clinical data

Fintech / Trading Compliance

Path: scenarios/fintech/ | Framework: LangChain + LangGraph | Compliance: SOX, MiFID II, PCI-DSS

A trading compliance agent bounded by 7 rulesets covering trade limits, account access, and data protection.

Rulesets

Key patterns demonstrated

• Claims-based authorization: principal.claims.trade_approval: true gates high-value trades without changing the user's role. See Principals for how claims work.

when:
  all:
    - args.quantity: { gt: 1000 }
    - principal.claims.trade_approval: { not_equals: true }

• Pattern-based blocking: Account IDs containing FROZEN or RESTRICTED markers are blocked via contains_any
• Quantitative thresholds: Numeric comparison operators (gt) for trade size limits
• Tiered roles: analyst < senior_trader < compliance_officer < risk_manager

Running it

cd scenarios/fintech
python fintech_agent.py --role analyst          # blocked large trades
python fintech_agent.py --role senior_trader    # allowed with trade_approval claim

Customer Support

Path: scenarios/customer-support/ | Framework: LangChain + LangGraph | Compliance: GDPR, PCI-DSS

A support-ticket agent bounded by 7 rulesets covering access control, refund limits, escalation rules, and PII redaction.

Rulesets

Key patterns demonstrated

• Financial thresholds: Refunds above $500 gated by role
• Operational workflow enforcement: Escalations require a reason field (not just RBAC)

when:
  any:
    - args.reason: { exists: false }
    - args.reason: { equals: "" }

• Data minimization: PII detection includes partial credit card numbers (****1234 pattern)
• Ticket lifecycle control: Only authorized roles can transition tickets to closed state

Running it

cd scenarios/customer-support
python support_agent.py --role support_agent     # limited access
python support_agent.py --role supervisor       # full access

DevOps / File Organizer

Path: scenarios/devops/ | Framework: OpenAI SDK (via OpenRouter) | Compliance: SOC 2

A file-organizer agent bounded by Python-based rulesets (not YAML). This scenario demonstrates the decorator rule API.

Rulesets (Python)

Key patterns demonstrated

• Python decorator API: Rulesets defined with @precondition, @postcondition, and @session_contract decorators instead of YAML
• Built-in rule: deny_sensitive_reads() provides instant secret-file protection with no configuration
• Factory pattern: Parameterized preconditions for reuse across different base paths

from edictum import precondition, Decision

@precondition("Bash")
def no_destructive_commands(envelope):
    cmd = envelope.bash_command
    if any(d in cmd.split() for d in ["rm", "rmdir", "shred"]):
        return Decision.fail(f"Destructive command blocked: {cmd}")
    return Decision.pass_()

• Self-correction: When the agent is blocked, it reads the block message and adjusts its approach (e.g., uses cp instead of mv to a restricted target)
• Compared with unguarded baseline: demo_without.py shows the same agent reading .env files and running rm -rf without enforcement

Running it

cd scenarios/devops
python demo_with.py      # enforced: destructive commands blocked
python demo_without.py   # unguarded: reads .env, runs rm -rf

When to use YAML vs Python rulesets

The pharma, fintech, and customer-support scenarios use YAML rulesets. The DevOps scenario uses Python decorators. Both approaches enforce identically -- the choice depends on your workflow:

Most teams start with YAML and add Python rulesets only for logic that cannot be expressed with built-in operators.

Running all scenarios

All scenarios support a --role flag and an observe mode flag. The observe mode flag varies between scenarios:

python <scenario>.py --role <role>       # set the principal role

# DevOps and Pharma scenarios:
python <scenario>.py --observe           # log but don't block

# Fintech and Customer-Support scenarios:
python <scenario>.py --mode observe      # log but don't block

See the edictum-demo README for setup instructions and environment variables.