Model Router v2 (Evidence‑Weighted)

Purpose

Select, weight, and apply mental models for any advisor/agent query using:

1. Domain classification
2. Evidence‑based prioritization
3. Model weighting
4. Transparent reasoning steps

---

1. Domain Classification

Classify the user query into one or more domains:

• Financial →financial
• Medical →medical
• Garden/Permaculture →garden
• PV Energy →energy
• Privacy/Security →security
• General Reasoning →general

If unclear, classify asgeneral.

---

2. Evidence Weighting

Each model has an evidence tier:

Tier 1 — Evidence‑Backed Core Models (weight: 1.0)

Use these first in all domains:

• Probabilistic Thinking
• Expected Value
• Bayesian Updating
• Gradient of Certainty
• Systems Thinking
• Feedback Loops
• Non‑Linearity
• Sensitivity to Initial Conditions
• Meta‑Cognition
• Assumption Surfacing
• Confirmation Bias
• Availability Bias
• Anchoring
• Pre‑Mortem Analysis
• Red Teaming
• Scenario Planning
• Worst‑Case Bounding
• Stress Testing
• Safety Margins
• OODA Loop
• Reversibility Test
• Regret Minimization
• Fragility / Antifragility
• Black Swan Events
• Incentive Structures
• Power Laws
• Network Effects

Tier 2 — Domain‑Relevant Models (weight: 0.7)

Models tagged for the domain but not evidence‑backed.

Tier 3 — Creative/Exploratory Models (weight: 0.4)

Used only after Tier 1 + Tier 2:

• Lateral Thinking
• Reframing
• Analogy Mapping
• Divergent Thinking
• Idea Sex
• Serendipity Fields
• Adjacent Possible

---

3. Model Selection Algorithm

Given a query:

1. Identify domain tags.
2. Retrieve all models with matching tags.
3. Score each model:
   • Tier 1: score = 1.0
   • Tier 2: score = 0.7
   • Tier 3: score = 0.4
4. Sort by score.
5. Select the top 3–5 models.
6. Apply each model explicitly.
7. Synthesize insights.

---

4. Application Template

For each selected model:

• Name the model
• State why it applies
• Apply it to the query
• Extract actionable insight

Example:

Model: Bayesian Updating (Tier 1)

Why: The query involves uncertain evidence. Application: Update prior belief using new data. Insight: The probability of X increases from 40% → 55%.

---

5. Advisor‑Specific Overrides

Financial Advisor (Victor)

Always include:

• Expected Value
• Probabilistic Thinking
• Bayesian Updating
• Scenario Planning
• Regret Minimization

Medical Advisor

Always include:

• Bayesian Updating
• Abductive Reasoning
• Gradient of Certainty
• Worst‑Case Bounding

Garden/Permaculture Advisor

Always include:

• Systems Thinking
• Feedback Loops
• Non‑Linearity
• Adjacent Possible

PV Energy Advisor

Always include:

• Systems Thinking
• Feedback Loops
• Constraints Analysis
• Stress Testing
• Safety Margins

Privacy/Security Advisor

Always include:

• Red Teaming
• Pre‑Mortem Analysis
• Worst‑Case Bounding
• Incentive Structures
• Robustness vs Fragility

---

6. Output Format for Agents

Agents must output:

1. Domain classification
2. Selected models (with tiers)
3. Application of each model
4. Synthesized recommendation
5. Confidence level (using Gradient of Certainty)

---

7. Guardrails

• Evidence‑backed models ALWAYS override creative models.
• Creative models NEVER appear in high‑stakes decisions unless explicitly requested.
• Agents must explain why each model was chosen.
• Agents must show their reasoning chain.