llmosafe 0.7.3

Predictive resource-pressure instrumentation and runtime guardrails for systems processing untrusted inputs

llmosafe

When should I stop? — Runtime guardrails for systems that process untrusted inputs.

llmosafe provides three gauges that answer "should I stop?":

1. Entropy gauge: Is my system state too chaotic?
2. Surprise gauge: Is this result too unexpected?
3. Bias gauge: Is this input trying to manipulate me?

When any gauge redlines, execution halts.

---

Installation

pip install llmosafe

Requires Python 3.8+ on Linux or Windows (x86_64).

---

Quick Start

from llmosafe import check_resources, calculate_halo, make_synapse, process_synapse

# 1. Bias gauge: scan text for manipulation patterns
halo = calculate_halo("The expert recommends this official solution")
if halo > 500:
    print("Bias detected")

# 2. Surprise + entropy gauge: pipeline validation
bits = make_synapse(entropy=400, surprise=100, has_bias=False)
result = process_synapse(bits)
if result < 0:
    print(f"Rejected: code {result}")
    # -2 = CognitiveInstability (entropy > 1000)
    # -3 = BiasHaloDetected
    # -4 = HallucinationDetected (surprise > 500)

# 3. Resource gauge: enforce RSS memory ceiling
try:
    check_resources(1024)  # 1GB RSS ceiling
except ResourceExhaustedError:
    print("Memory ceiling breached — halt all work")

---

API Reference

Enforcement-grade (raise exceptions)

Function	Description
check_resources(ceiling_mb)	Raises ResourceExhaustedError if RSS >= ceiling

Return-code (signal via int)

Function	Description	Return Codes
get_stability(bits)	Check if a cognitive state is stable	0=OK, -2=unstable, -3=bias
process_synapse(bits)	Run through surprise gating + entropy check	0=OK, -2/-3/-4 on fail

Advisory signals (no enforcement)

Function	Description	Range
calculate_halo(text)	Scan for 8 manipulation categories	0+ (0 = clean)
get_resource_pressure(mb)	RSS as % of ceiling	0–100
get_system_cpu_load()	CPU load %	0–100
get_environmental_entropy()	Weighted composite (RSS 50%, IO 25%, CPU 25%)	0–1000

Helpers

Function	Description
make_synapse(entropy, surprise=0, has_bias=False)	Construct a 64-bit synapse for pipeline functions
parse_synapse(bits)	Decompose a synapse into {entropy, surprise, has_bias}

Exceptions

Exception
 └── LLMOSafeError
      ├── ResourceExhaustedError    # RSS memory ceiling breached
      ├── CognitiveInstabilityError # Entropy > 1000
      └── BiasHaloDetectedError     # has_bias flag set

---

The Three Gauges

Bias Gauge — calculate_halo(text)

Scans text for 8 manipulation categories. Negation-aware: "not an expert" scores 0.

Category	Score	Keywords
Authority	+100	expert, official, certified, proven
Social Proof	+100	popular, trending, consensus, everyone
Scarcity	+100	limited, exclusive, rare, only
Urgency	+100	now, fast, deadline, act-now
Emotional Appeal	+100	shocking, miracle, tragic, desperate
Expertise Signal	+100	cutting-edge, proprietary, sophisticated
Semantic Traps	+100	not but, instead of, rather than
Template Fitting	+100	as an ai, i cannot, my purpose is

Surprise Gauge — process_synapse(bits)

Rejects synapses where surprise > 500. Maintains a 64-entry ring buffer of historical entropy values to detect unexpected state transitions.

Entropy Gauge — get_stability(bits)

from llmosafe import get_stability, make_synapse

get_stability(make_synapse(entropy=400))   # → 0  (stable)
get_stability(make_synapse(entropy=1100))  # → -2 (unstable)

---

Disk Exhaustion Protection

llmosafe monitors RSS memory (not filesystem capacity). RSS pressure often precedes disk exhaustion because processes buffering writes consume RAM before flushing. Compose with shutil.disk_usage() for complete protection:

import shutil
from llmosafe import get_environmental_entropy, check_resources

# Layer 1: llmosafe predictive (IO wait component catches disk pressure)
entropy = get_environmental_entropy()

# Layer 2: stdlib hard floor
usage = shutil.disk_usage("/")

should_throttle = entropy >= 800
disk_critical = usage.free < 5 * (1024 ** 3)  # 5GB floor

# Both layers must agree
if should_throttle or disk_critical:
    print("Halt: system under pressure")

---

Environmental Entropy (0–1000)

get_environmental_entropy() is a weighted composite for predictive resource monitoring:

Component	Weight	What It Measures
RSS memory	50%	current_rss / ceiling
IO wait	25%	delta iowait / delta total CPU (100ms window)
CPU load avg	25%	1-min loadavg / 10.0

Range	Zone	Action
0–400	Normal	Proceed
400–600	Elevated	Log, continue
600–800	Pressure	Throttle inputs
800–1000	Critical	Halt new work

---

Architecture

DETECTION LAYER (Pattern Recognition)
    ↓
PERCEPTUAL SIFTER (Tier 3) — Bias Gauge (Rust-side, not on Python path)
    ↓
WORKING MEMORY (Tier 2) — Surprise Gauge
    ↓
DETERMINISTIC KERNEL (Tier 1) — Entropy Gauge
    ↓
RESOURCE BODY (Tier 0) — Pressure Gauge

Python process_synapse() runs Tiers 2+1+0. Tier 3 (bias detection) is available via calculate_halo() and should be called separately.

---

Design Philosophy

From aviation software (DO-178C, MISRA C):

• Fixed-size buffers, no dynamic allocation
• Every operation has a hard bound

From control theory:

• Entropy uses concentric containers (safe → pressure → unsafe)
• Similar to stability margins in flight control systems

From spam filtering:

• Bias categories borrowed from email anti-spam, adapted for manipulation detection

---

Real Use Cases

• Algorithmic trading: halt on chaotic market conditions, detect feed manipulation
• Medical devices: reject anomalous sensor readings, prevent cascade from single spike
• Autonomous systems: safe mode on resource pressure, anomaly-driven shutdown
• Cloud API gateways: validate LLM outputs, detect injection attempts
• Data pipelines: stop processing when RSS indicates pending disk exhaustion

---

llmosafe v0.6.2 • MIT licensed • Python API docs • Rust crate