Research Tool · Recursive Creation Model · Free

Know Your AI's Risk Score
Before Your Board Does

Risk Management applies a mathematically grounded 5-vector risk framework to your AI architecture — surfacing thermodynamic bottlenecks, autonomy runaway, and infrastructure gaps in a free 10-minute conversation.

See the Framework →
Trusted for VC Due Diligence Board Risk Reports CTO Architecture Reviews
The Science Behind the Score

The Recursive Creation Model

Intelligence at every scale creates a child on an independent substrate to avoid resource competition with its creator. This single axiom produces a measurable, predictive risk structure for any AI system.

L1
God / Universe
Quantum vacuum → matter → physics
→
L2
Human / Carbon
Biological neurons · 20 watts
→
L3
AI / Silicon
Data centres · Gigawatts · You are here
→
L4
Quantum / Nano-Bio
Phase transition · Landauer minimum
The Imbalance Penalty — δ (delta)
δ = Mean(Ns, Au) − Mean(Ee, Cb, Sp)
δ > 0  →  Risk = 1 − e−4.2 · δ1.45  [ambition outpaces substrate → exponential risk]
δ ≤ 0  →  Risk = 0.10 · e2.8δ  [infrastructure surplus → near-zero risk]
⚡
Ee
Energy Efficiency
Compute-to-watt ratio, quantisation, inference optimisation
🔗
Cb
Comm Bandwidth
Inter-agent throughput, messaging architecture, latency headroom
🧠
Sp
Storage Persistence
Memory durability, GraphRAG density, crash-safe knowledge
🔀
Ns
Node Specialisation
Agent role diversity, modularity, coupling graph structure
🤖
Au
System Autonomy
Autonomous decision scope, loop guards, human-in-the-loop
Verified in code

ML risk engine — production verification targets

The AP003 scorer trains on 200 synthetic enterprise architecture profiles and validates on hold-out data before deployment. Thresholds are enforced in ap003/risk_engine.py run_verification().

200
Training profiles
35% balanced · 25% underutilised · 40% imbalanced
≤0.05
Hold-out MAE
GradientBoostingRegressor on RCM risk score
≥0.95
Hold-out R²
Non-linear Imbalance Penalty δ mapping
15%
Anomaly detection
IsolationForest flags runaway autonomy outliers

Five bottleneck classes — with engineering mitigations

Energy Starvation

INT8/FP8 quantisation recovers 30–45% wasted inference energy on large models.

Bandwidth Saturation

10 × 50 ms HTTP agent chains = 500 ms before work. Queues / gRPC cut 60–80%.

Memory Persistence Collapse

GraphRAG delivers 40–60% better grounding vs flat vector top-k on knowledge tasks.

Node Hyper-Specialisation

12–15+ specialists without a broker → long-tail routing failures.

Autonomy Runaway

Autonomy >0.8 without depth guards is the highest-risk RCM configuration.

Engineering signals sourced from the AP003 architectural knowledge base (ap003/rag_context.py) — the same probes power the conversational auditor.

Why This Matters

RCM vs Standard AI Predictions

Standard AI forecasts are human-centric — they ask "how will AI affect society?" The RCM asks "what are the physical constraints on any intelligence substrate?" This is a fundamentally different, and more rigorous, starting point.

Standard AI (The Hype)
RCM Prediction (The Architecture)
Primary Horizon
Software & Agents — AGI, smarter LLMs, autonomous workers
Substrate Phase Shift — away from silicon & the electrical grid
Relation to Humans
Competitor or tool — "Will AI take our jobs?"
Non-competitor — the creator–child substrate independence axiom
Level 4 Form
Bigger data centres, more GPUs, global software networks
Quantum matter or nano-biological hybrid systems
Core Bottleneck
Data & training — more data, better alignment algorithms
Thermodynamic limit — Landauer's Principle, the 20-watt gap
Framing
Human-centric: society, jobs, language, culture
System-centric: physics, thermodynamics, substrate capacity
The Process

How the Audit Works

01
Conversational Extraction
Describe your AI system naturally. The agent extracts 5 vector scores through targeted questions — no spreadsheets, no forms.
02
ML Risk Engine
A GradientBoosting + IsolationForest model trained on 200 enterprise architectures scores your system against the RCM Imbalance Penalty.
03
Bottleneck Diagnosis
Receive a risk tier (LOW / MODERATE / HIGH / CRITICAL), primary bottleneck identification, gradient sensitivity analysis, and an engineering mitigation playbook.
Built on Insight IT Libraries

Powered by Our Own Stack

Risk Management runs entirely on proprietary libraries built and maintained by Insight IT Solutions — no third-party AI infrastructure dependencies. Each library handles a distinct layer of the audit pipeline.

PrismLang
PrismLang

Intent Classification & Vector Encoding

Every agent turn is encoded via a @prism_node decorator using a 7-category RCM taxonomy (the 5 audit vectors + Theory + Session). PrismLang's PrismProjector soft-classifies each message into the correct vector bucket before extraction — so ambiguous descriptions like "our GPUs are always on" land in Energy Efficiency, not general context. A BoundaryTranslator emits a full PrismLang audit summary at session end.

PrismLang product page →
PrismRAG
PrismRAG

Domain Remapping & Semantic Retrieval

Two roles. First, before the ML risk engine scores your vectors, a 40-rule weighted mapping (e.g. "kafka" → Comm_Bandwidth ×1.0, "graphrag" → Storage_Persistence ×1.0) nudges raw extracted values toward the correct RCM bucket using your full conversation as signal. Second, knowledge base retrieval uses Gemini embedding cosine-similarity ranked by PrismRAG instead of keyword overlap — surfacing the right engineering context even for non-standard vocabulary.

PrismRAG product page →
PrismResonance
PrismResonance

Wave-Interference Theory Retrieval

The five RCM theory entries are embedded and ingested into a PrismResonance sidecar. When you ask a framework question ("what is Level 4?", "explain the delta formula"), retrieval uses biologically-inspired wave-interference scoring rather than static cosine similarity. Entries that are co-cited in the same session record a Hebbian co-activation signal — over time, related theory entries phase-align and surface together, producing progressively more coherent multi-topic answers.

PrismResonance product page →
All three libraries degrade gracefully — when a licence key or API key is absent, each falls back to the next tier (semantic → keyword) without disrupting the audit.
Start Now

Talk to the Auditor

Ask about the future of AI, compare RCM predictions with Gemini or ChatGPT, or describe your AI architecture to receive a live risk score.

◎

Free Research Tool

No access key required. Start a conversational architecture audit and receive a live RCM risk score with bottleneck diagnosis.

Who Uses Risk Management

Built for High-Stakes AI Decisions

💼

VC Due Diligence

Quantify AI infrastructure risk before term sheets. Identify whether a portfolio company's architecture can sustain scale without thermodynamic collapse.

🏛️

Board Risk Reports

Translate complex AI architecture decisions into a single risk score and bottleneck narrative that non-technical board members can act on.

⚙️

CTO Architecture Reviews

Run the 5-vector audit before major scaling decisions. The gradient sensitivity analysis shows exactly which investment moves the risk needle most.

🔬

AI Research Positioning

The RCM framework is falsifiable and mathematically structured — a rigorous alternative to anthropomorphic AI forecasting for academic and commercial research.