Skip to content

Decision Kernel Lite — Choosing Under Uncertainty

Live demo (embedded)

What you can do in the demo (30 seconds)

  1. Define Scenarios + Probabilities (they’ll be normalized if they don’t sum to 1).
  2. Define Actions + Losses (loss matrix = action × scenario).
  3. Choose a decision lens:
  4. Expected Loss
  5. Minimax Regret
  6. CVaR (tail-risk)
  7. Copy the generated Decision Card (exec-ready rationale).

What this proves

  • Decisions fail because uncertainty is handled informally: probabilities are debated, downside is underestimated, and justification is retrospective.
  • You don’t need certainty to decide. You need explicit assumptions, bounded downside, and defensible trade-offs.
  • Different lenses are not “better/worse” — they represent different risk postures.

Core concept

A decision is defined by four primitives:

Actions × Scenarios × Probabilities × Losses

The kernel evaluates each action under three lenses and outputs one recommended action + justifications.

Decision lenses (when to use what)

1) Expected Loss (risk-neutral)

Use when:

  • decisions repeat frequently
  • probabilities are reasonably trusted
  • variance is acceptable

Optimizes:

  • long-run average pain

2) Minimax Regret (robust / political safety)

Use when:

  • probabilities are unreliable or contested
  • it’s one-shot or high-accountability
  • post-hoc defensibility matters

Optimizes:

  • “what will I regret least in hindsight?”

3) CVaR (tail-risk protection)

Use when:

  • rare bad outcomes are unacceptable (ruin / safety / bankruptcy)
  • downside is asymmetric and must be bounded
  • survival > average performance

Optimizes:

  • average loss in the worst cases (tail), not the overall average

Decision logic (high level)

flowchart TD
  A[Inputs: Actions, Scenarios, Probabilities, Losses] --> B[Compute Expected Loss per action]
  A --> C[Compute Regret matrix per action × scenario]
  A --> D[Compute CVaR per action at alpha]
  B --> E{Primary rule}
  C --> E
  D --> E
  E -->|Expected Loss| F[Choose argmin Expected Loss]
  E -->|Minimax Regret| G[Choose argmin Max Regret]
  E -->|CVaR| H[Choose argmin CVaR]
  F --> I[Decision Card]
  G --> I
  H --> I

Rule recommendation (simple heuristic)

The app includes a transparent heuristic:

  • if tail risk dominates average risk → recommend CVaR
  • otherwise → recommend Expected Loss

This is advisory only; you can override it. Governance is preserved.

Outputs you get

  • One recommended action

  • Evidence table with:

  • Expected Loss

  • Max Regret
  • CVaR@alpha
  • Regret table (action × scenario)
  • A copy/paste Decision Card (for a memo or exec deck)

Downloads

Slides