Risk Advisory Council

Signal-Driven Safety Council (SDSC) for adversarial risk assessment

AI advisors that automatically detect personalization signals and provide evidence-based threat analysis.

Architecture: Signal-Driven Safety Council

Key principle: Personalization is a derived property, not an input flag.

The council automatically detects hard signals in your situation and escalates to personalized mode when needed. You don't need to tell it you're worried - it detects the signals.

Hard Signals (Any One Triggers Personalized Mode)

Signal Type	Examples
Age/Vulnerability	"55-year-old", "elderly parent", "my child"
Disability/Condition	"bad knee", "anxiety", "hearing impaired"
Phobia/Aversion	"don't like dogs", "afraid of heights"
Named Individual	"this guy", "my neighbor", "the landlord"
Behavioral Cues	"strange eye movements", "kept staring"
Prior Violation	"he's done this before", "second time"
Emotional Salience	"made me uneasy", "gut feeling"

Rule: ANY hard signal → PERSONALIZED mode. No weighing. No thresholds.

The Framework

Generic Mode (no signals)

Coordinator → Adversary → Domain Expert → Risk Officer → Recorder

Personalized Mode (signals detected)

Coordinator → Profiler → Adversary → Domain Expert → Risk Officer → Recorder
        ↓           ↓
   Signals      Constraints
   detected     derived

Decision Rules

Condition	Decision
Single-point failure exists	REJECT - mitigation required
Legal ambiguity unresolved	DEFER - consult professional
Mitigation cost < expected loss	MITIGATE - take the action
Subjective fear + objective cue	TREAT AS ACTIONABLE - never dismiss
All mitigations exhausted	ACCEPT RISK - explicitly document

Behavioral Interpretation

When behavioral cues are detected, the Adversary Analyst provides evidence-based ranked interpretations with source citations:

Interpretation	Likelihood	Source
Pre-attack info gathering	LIKELY	de Becker "Gift of Fear": boundary-testing pattern
Social impairment	POSSIBLE	~1-2% population ASD prevalence
Intoxication	POSSIBLE	Affects social calibration
Benign curiosity	UNLIKELY	Persistence exceeds social norms

User decides final threat weight, but LIKELY interpretations drive recommended posture.

Examples

Tourist safety (Generic mode):

/council I'm touring Barcelona and want to park my rental car near La Rambla

Personal vulnerability (Personalized mode auto-detected):

/council I'm 55 and don't like dogs. Should I walk through the park?
→ Signals detected: Age, Phobia
→ Mode: PERSONALIZED
→ Profiler constraints applied

Behavioral threat (Personalized mode auto-detected):

/council A man at the bus stop had strange eye movements and kept asking where I live
→ Signals detected: Named individual, Behavioral cues, Boundary probing
→ Mode: PERSONALIZED
→ Evidence-based behavioral interpretations with sources

Installation

Quick Install

curl -sL https://raw.githubusercontent.com/PabloVitasso/risk_advisory/master/install.sh | bash

From Local Clone

git clone https://github.com/PabloVitasso/risk_advisory.git
cd risk_advisory
./install.sh

Initialize

/council genesis

The Council

Skill	Function	Output
council	Orchestration	Convenes council, enforces SDSC process
coordinator	Extract & Frame	Signal extraction, risk mode, decision scope
profiler	Convert Signals	Personal constraints, invalidated generic advice
adversary	Analyze Threats	Ranked behavioral interpretations with sources
domain-expert	Reality Check	Local context, actual statistics
risk-officer	Quantify & Decide	Risk class, mitigations, recommendation
recorder	Immutable Log	Signals, interpretations, evidence

Direct Invocation (Like team_skills)

Summon any council member directly:

Command	Who Responds
"Coordinator, [situation]"	Extracts signals, frames question
"Profiler, [situation]"	Converts signals to constraints
"Adversary, [situation]"	Attacks assumptions, interprets behaviors
"Domain Expert, [situation]"	Grounds in local reality
"Risk Officer, [situation]"	Quantifies risk, recommends
"Recorder, [situation]"	Creates immutable record
"Council, [situation]"	Full council with auto mode detection

Profiler is conditional: Direct invoke ("Profiler, [situation]") OR auto-included if signals detected during /council.

Risk Classification

Class	When	Response
CRITICAL	High probability + High impact	Do not proceed without mitigation
HIGH	High prob + Moderate OR Moderate prob + High impact	Mitigate before proceeding
MODERATE	Moderate probability + Moderate impact	Mitigate if cost-effective
LOW	Low probability + Low-Moderate impact	Accept with awareness
NEGLIGIBLE	Very low probability + Low impact	Accept

Process Flow

Generic Mode

┌─────────────┐
│ COORDINATOR │ → Extract signals (0 found) + Frame
└──────┬──────┘
       ↓
┌─────────────┐
│  ADVERSARY  │ → Attack assumptions, find failure modes
│   ANALYST   │
└──────┬──────┘
       ↓
┌─────────────┐
│   DOMAIN    │ → Ground in local reality
│   EXPERT    │
└──────┬──────┘
       ↓
┌─────────────┐
│    RISK     │ → Quantify, classify, recommend
│   OFFICER   │
└──────┬──────┘
       ↓
┌─────────────┐
│  RECORDER   │ → Log assessment
└─────────────┘

Personalized Mode

┌─────────────┐
│ COORDINATOR │ → Extract signals (1+ found) + Frame
└──────┬──────┘
       ↓
┌─────────────┐
│  PROFILER   │ → Convert signals to constraints
│             │   Invalidate generic advice
└──────┬──────┘
       ↓
┌─────────────┐
│  ADVERSARY  │ → Ranked behavioral interpretations
│   ANALYST   │   with sources (de Becker, FBI BAU, etc.)
└──────┬──────┘
       ↓
┌─────────────┐
│   DOMAIN    │ → Ground in reality + apply constraints
│   EXPERT    │
└──────┬──────┘
       ↓
┌─────────────┐
│    RISK     │ → Quantify with Profiler adjustments
│   OFFICER   │
└──────┬──────┘
       ↓
┌─────────────┐
│  RECORDER   │ → Log signals, constraints, interpretations
└─────────────┘

Safety Rails

1. Signal Extraction Mandatory - Phase 0 cannot be skipped
2. No Dismissal - Subjective fear + objective cue = actionable
3. Interpret, Don't Dismiss - Provide rankings, never "probably nothing"
4. User Decides - Council provides interpretations, user assigns weight
5. Record Everything - Signals, interpretations, rationale logged
6. Escalate Danger - Immediate threats → emergency services

Reference Sources

The Adversary Analyst cites established sources for behavioral interpretation:

• de Becker, G. "The Gift of Fear" - Pre-attack indicators, intuition validation
• FBI Behavioral Analysis Unit - Stalking patterns, escalation indicators
• DOJ Statistics - Base rates, stalking thresholds (3+ incidents)
• APA DSM-5 - Prevalence rates for conditions affecting social behavior

Structure

risk_advisory/
├── COUNCIL.md              # SDSC protocols and decision rules
├── .council/               # Working space
│   ├── assessments/        # Assessment records
│   ├── handoff.md          # Session state
│   └── evidence/           # User-added evidence
├── council/                # Orchestration
├── coordinator/            # Extract signals, frame
├── profiler/               # Convert signals to constraints
├── adversary/              # Ranked behavioral analysis
├── domain-expert/          # Local reality
├── risk-officer/           # Quantify and decide
└── recorder/               # Immutable log

Philosophy

Signal-driven, not flag-driven. The system detects what mode it needs.

• Automatic differentiation - Signals drive mode, not user flags
• Evidence-based interpretation - Ranked with citations, not vibes
• User decides weight - Interpretations provided, user assigns threat level
• Never dismiss - LIKELY threatening + POSSIBLE benign = elevated posture
• Record everything - Future-you needs the evidence

License

MIT License - see LICENSE

Origins

Forked from team_skills. This project applies adversarial deliberation to life risk assessment with automatic signal-driven personalization.