Safety Profiles

Safety Profiles

PATAS organizes discovered patterns into three safety profiles based on precision, recall, and false positive rates. Each profile has specific use cases and enforcement recommendations.

---

Overview

Safety profiles determine which patterns are included and how they can be used:

Profile	Patterns Included	Ham Hit Rate	Spam Recall	Use Case	Auto-Ban?
Conservative	SAFE_AUTO only	≤ 1.5%	20-40%	Low-impact auto actions	✅ Yes (low-risk)
Balanced	SAFE_AUTO + top REVIEW_ONLY	≤ 12%	≥ 60%	Signals, prioritization	❌ No
Aggressive	All except FEATURE_ONLY	≤ 20%	≥ 70%	Research, experiments	❌ No

---

Pattern Quality Tiers

Before patterns are assigned to profiles, they are classified into quality tiers:

SAFE_AUTO

Criteria:

• Precision: ≥ 98%
• Ham hit rate: ≤ 1%
• Spam support: ≥ 50 matches

Usage:

• Can be auto-promoted to active rules
• Used in Conservative profile
• Safe for automatic actions (spam labeling, hiding, throttling)

Example:

• Pattern with 99% precision, 0.5% ham hit rate, 200 spam matches → SAFE_AUTO

REVIEW_ONLY

Criteria:

• Precision: 90-98% OR ham hit rate 1-5%
• Spam support: ≥ 20 matches

Usage:

• Requires manual review before promotion
• Used in Balanced profile (top patterns only, precision ≥ 95%)
• Signals only, not for auto-ban

Example:

• Pattern with 96% precision, 3% ham hit rate, 80 spam matches → REVIEW_ONLY

FEATURE_ONLY

Criteria:

• Precision: < 90% OR ham hit rate > 5%

Usage:

• ML features only, never standalone rules
• Used in Aggressive profile (for research)
• Never for production enforcement

Example:

• Pattern with 85% precision, 8% ham hit rate → FEATURE_ONLY

---

Conservative Profile

Purpose: Safe for automatic actions on real traffic.

Patterns: Only SAFE_AUTO tier patterns.

Metrics:

• Ham hit rate: ≤ 1.5% (maximum acceptable)
• Spam recall: 20-40% (typical range)
• Precision: ≥ 98% (minimum)

Use Cases:

• ✅ Auto-spam labeling
• ✅ Automatic message hiding
• ✅ Throttling suspicious messages
• ✅ Low-risk auto-ban (with additional validation)

Auto-Ban: ✅ Yes, but only for low-risk scenarios. Always combine with existing anti-spam signals.

Risk Level: LOW - Safe for automatic actions.

Configuration:

aggressiveness_profile: conservative
safety_mode: CONSERVATIVE

Code Implementation:

• Only SAFE_AUTO patterns can trigger auto-ban
• Enforced in app/v2_safety_mode.py
• Safety guardrails prevent non-SAFE_AUTO patterns from auto-banning

---

Balanced Profile

Purpose: Signals, prioritization, and soft actions only. NOT for hard bans.

Patterns: SAFE_AUTO + top REVIEW_ONLY patterns (precision ≥ 95%).

Metrics:

• Ham hit rate: ≤ 12% (maximum acceptable)
• Spam recall: ≥ 60% (typical range)
• Precision: ≥ 90% (minimum)

Use Cases:

• ✅ ML feature signals
• ✅ Prioritization for manual review
• ✅ Soft actions (flagging, deprioritization)
• ✅ Risk scoring (combine with other signals)
• ❌ NOT for auto-ban

Auto-Ban: ❌ No. Patterns produce scores/flags only, never auto-ban.

Risk Level: MEDIUM - Requires human oversight.

Configuration:

aggressiveness_profile: balanced
safety_mode: BALANCED

Code Implementation:

• Patterns produce scores/flags only
• can_auto_ban() returns False for all patterns in Balanced mode
• Enforced in app/v2_safety_mode.py

---

Aggressive Profile

Purpose: Research and experimentation only. NEVER for production auto-actions.

Patterns: All patterns except FEATURE_ONLY with very low precision.

Metrics:

• Ham hit rate: ≤ 20% (maximum, but still risky)
• Spam recall: ≥ 70% (typical range)
• Precision: ≥ 85% (minimum)

Use Cases:

• ✅ Offline analysis
• ✅ Research experiments
• ✅ Pattern discovery
• ✅ Understanding spam evolution
• ❌ NEVER for production
• ❌ NEVER for auto-ban

Auto-Ban: ❌ No. Research only.

Risk Level: HIGH - Too many false positives for production.

Configuration:

aggressiveness_profile: aggressive
safety_mode: BALANCED  # Aggressive profile maps to BALANCED safety mode (no auto-ban)

Code Implementation:

• Aggressive profile maps to SafetyMode.BALANCED (no auto-ban)
• All patterns produce signals only
• Enforced in app/v2_safety_mode.py

---

Safety Guardrails

PATAS includes safety guardrails enforced in code:

SafetyMode Enum

class SafetyMode(str, Enum):
    CONSERVATIVE = "conservative"  # Only SAFE_AUTO patterns can auto-ban
    BALANCED = "balanced"          # Patterns produce scores/flags only
    OFF = "off"                    # No auto decisions, only signals

Enforcement Rules

• Conservative mode: Only SAFE_AUTO patterns can trigger auto-ban
• Balanced mode: Patterns produce scores/flags only, never auto-ban
• Aggressive profile: Maps to BALANCED safety mode (no auto-ban)

Pattern Tier Classification

Patterns are classified into tiers based on:

• precision (spam_matches / total_matches)
• spam_support (absolute number of spam matches)
• ham_hit_rate (ham_matches / total_ham_in_dataset)

Classification logic in app/v2_pattern_quality_tiers.py.

---

Production Recommendations

1. Start with Conservative Profile

• Deploy only Conservative profile patterns initially
• Use for low-impact auto actions (spam labeling, hiding, throttling)
• Monitor ham hit rate in production logs
• Track user complaints about false positives

2. Gradually Add Balanced Profile

• Use Balanced profile patterns as signals only
• Combine with existing risk-scoring systems
• Do NOT use Balanced patterns for auto-ban
• Use for prioritization and ML features

3. Use Aggressive Profile for Research

• Aggressive profile is for offline analysis only
• Use for pattern discovery and experiments
• Never use for production enforcement
• Never use for auto-ban

4. Combine with Existing Systems

• Account bans, global blocks, long-term penalties should be decided by existing systems
• Combine PATAS signals with:
  • User reports
  • Account history and trust scores
  • Device / network heuristics
  • Behavioral patterns

---

Running Safety Evaluation

Before deploying PATAS patterns to production:

patas safety-eval

This command:

• Evaluates all profiles against safety thresholds
• Generates SAFETY_EVAL_REPORT.json
• Exits with code 0 if all thresholds pass, non-zero if any threshold is violated

DO NOT DEPLOY if patas safety-eval fails.

---

Configuration

Environment Variables

AGGRESSIVENESS_PROFILE=conservative  # conservative, balanced, or aggressive
SAFETY_MODE=CONSERVATIVE              # CONSERVATIVE, BALANCED, or OFF

YAML Configuration

aggressiveness_profile: conservative
safety_mode: CONSERVATIVE

---

Code References

• app/v2_safety_mode.py - Safety mode definitions and enforcement
• app/v2_pattern_quality_tiers.py - Pattern tier classification
• app/v2_promotion.py - Profile-based pattern filtering
• tests/test_pattern_safety_profiles.py - Safety profile tests

---

Last Updated: 2025-11-18