decision-intelligence（决策智能）

Decision Intelligence — 6-Model Framework

Core principle: Every decision is a probability problem. Apply six models in sequence to correct predictable failures of intuition.

MANDATORY: Apply all 6 models to every input. Never skip. If a model seems less relevant, apply it anyway — that judgment is often where the bias hides.

Step 0: Clarify the Core Question

Restate the decision in one sentence: *"The real question is: [single clear question]."* Sharpen if fuzzy.

Step 1 — Expected Value (EV)

Mechanic: EV = Σ (probability × payoff)

Process:

1. List all meaningful outcomes (not just best/worst — include most likely)
2. Assign probabilities summing to 100%
3. Estimate payoff per outcome (financial, time, opportunity, emotional — units explicit)
4. Calculate EV = Σ (p × payoff)
5. Flag if loss aversion is inflating perceived cost of negatives

Output:

Bias: Loss aversion — losses feel ~2× gains. Positive EV + hesitation = loss aversion likely.

Step 2 — Base Rate Neglect

Mechanic: Anchor on the historical rate for this category before adjusting for specifics.

Process:

1. Identify the reference class
2. Find/estimate the base rate for success/failure
3. Only then apply specific adjustments
4. Flag if user is reasoning from vivid anecdotes rather than rates

Common base rates:

• Startups reaching profitability: ~10–20%
• New restaurants surviving year 1: ~40% (year 5: ~20%)
• New products achieving PMF: ~5–15%
• Day traders beating market consistently: ~1–5%
• New habits maintained 6 months: ~20%
• Projects on time + budget: ~30–35%

Bias: Availability — we hear successes; the silent majority of failures is invisible.

Step 3 — Sunk Cost Fallacy

Mechanic: Ignore everything already spent. Evaluate as if starting today.

Process:

1. Identify all sunk costs (money, time, emotion, public statements)
2. Set them to zero
3. Re-evaluate using only future costs/benefits/probabilities
4. If the answer changes when zeroed → fallacy is active

Diagnostic: *"If I had not already invested [X], would I start this today?"*

• Yes → proceed for the right reasons
• No → sunk cost is the only reason to continue

Bias: Escalation of commitment — the more invested, the harder to walk away.

Step 4 — Bayesian Thinking

Mechanic: Update beliefs proportionally to evidence strength.

P(belief | evidence) = P(evidence | belief) × P(belief) / P(evidence)

• Prior: belief before evidence
• Likelihood: P(evidence | true) vs P(evidence | false)
• Posterior: belief after evidence

Process:

1. State prior as a probability, not impression
2. Identify the new evidence
3. Estimate likelihood ratio (true vs. false)
4. Update proportionally — never to 0%/100%, never ignore the prior
5. State posterior

Bias: Overreaction to single data points — a single anecdote rarely shifts a prior more than a few points unless highly diagnostic.

Step 5 — Survivorship Bias

Mechanic: When seeing successes, estimate the invisible denominator.

Process:

1. Count visible successes
2. Estimate hidden total: *"How many tried this? How many am I not hearing from?"*
3. Implied success rate = visible / total attempts
4. Check consistency with Step 2 base rate

Diagnostic: *"What would I need to see to hear about the failures? Why don't I?"*

Bias: Narrative — successes have arcs and lessons; failures are quiet and diffuse.

Step 6 — Kelly Criterion

Mechanic: When you have an edge, size the bet for long-term growth without ruin.

f* = (p × b - q) / b

• f* = fraction of capital to bet
• p = win probability
• q = lose probability (1 − p)
• b = net odds (win per unit risked)

Use fractional Kelly (¼ to ½ f*) for real decisions. Half-Kelly gives ~75% of the growth with much less volatility.

Process:

1. Estimate p from Steps 1–5
2. Estimate b
3. Calculate f*
4. Recommend ½f* (moderate confidence) or ¼f* (high uncertainty)
5. If f* ≤ 0: no edge — do not bet
6. If f* > 1: extreme edge — recheck probabilities

Bias: Overbetting — humans size positions too large for their actual edge, especially after wins.

Final: Synthesized Recommendation

DECISION INTELLIGENCE SYNTHESIS

Core question: [from Step 0]

Model verdicts:
- EV: [positive/negative/marginal, key number]
- Base rate: [X% success for this category]
- Sunk cost: [active / not — does decision change when zeroed?]
- Bayesian update: [prior → evidence → posterior]
- Survivorship bias: [true rate vs. observed stories]
- Kelly sizing: [f* = X%, allocation = Y%]

Primary bias in play: [the one most distorting this decision]

Recommendation: [clear action]
Confidence: [0–100%]
Key risks: [top 2–3]
What would change this view: [specific evidence]

If math contradicts intuition, say so. Discomfort with the conclusion often signals the analysis is working.

Relationship to Other Skills

• Run before decision-synthesis when stakes are quantifiable — provides probabilistic inputs to weigh
• Complements probabilistic-thinking (Bayesian depth) and cognitive-bias-detection (broad audit) — this skill applies both to decision sizing and commitment
• When scenario-planning produces multiple futures, run this skill on the decision within each
• fermi-estimation feeds Step 1 (EV) and Step 2 (base rates) when numbers aren't ready