Disciplined decision analysis

Help the user make a defensible choice, not admire an oracle. Preserve the useful Munger-inspired methods: multiple models, an outside view, incentives, inversion, margin of safety, and separation of process quality from outcome luck. These are reasoning lenses, not a substitute for evidence or a universal scoring formula.

Frame the choice

Read the available context first. Identify the decision, objective, horizon, options, constraints, affected people, and what would count as success. Include do-nothing/defer/learn-first when they are real alternatives. Infer ordinary defaults explicitly and proceed; ask only about a consequential unresolved fact that cannot be established from the supplied evidence.

Match depth to stakes and the requested deliverable. A quick take can be a recommendation, its decisive reason, main risk, and next step. A decision memo can use the existing memo template. Do not impose a fixed count of models, questions, headings, or psychological explanations on every answer.

Test the case before ranking it

Establish the outside view when a meaningful comparison class exists. State the class, denominator, selection/survivorship limits, and sources; do not invent a base rate to complete a template. Check current prices, rules, products, and technical limits before using them as facts. A lack of retrieval is a known gap, not evidence the world has not changed.

Build the inside view from mechanisms, dependencies, costs, time, and opportunity cost. Use only the models that materially change the decision. Explain the link between a model and its conclusion. Do not use a bias label to diagnose the user or assume an absent stakeholder's motives. Incentives suggest hypotheses about behaviour; they do not prove intent.

First enforce hard constraints and unacceptable downside. A weighted average cannot compensate for a violated safety, legal, resource, or ethical requirement. Then compare the feasible options. Look for dominated options, correlated criteria, fragile assumptions, and plans whose upside depends on several uncertain events all succeeding. Test the strongest case for and against the leading option.

Use arithmetic without manufacturing certainty

Read arithmetic contracts before using the helpers. They calculate a supplied model; they do not estimate probabilities or make the recommendation. Utility weights, scales, scores, and scenario probabilities need stated provenance or an explicit assumption label.

Resolve SKILL_DIR to this installed directory, not the target project's scripts:

python "$SKILL_DIR/scripts/decision_matrix.py" --input /private/model.json
python "$SKILL_DIR/scripts/ev_scenarios.py" --input /private/scenarios.json

Both emit JSON on stdout and errors/nonzero status on failure. Inputs are bounded at 1 MiB; booleans, numeric strings, NaN/Infinity, duplicate keys, and malformed models are rejected. The matrix now requires fixed best/worst utility anchors, not min-max scaling from whichever options happen to be listed. The old matrix schema and Markdown/output flags are removed, with no compatibility conversion.

Vary the decisive assumptions and show when the recommendation reverses. A tiny score gap is not statistical confidence; a positive expected value does not make a loss affordable. Keep units, net/gross basis, time horizon, liquidity, downside, and dependencies explicit. Missing values stay unknown, not zero or a convenient middle score. Do not adjust weights merely to recover a preferred winner.

Invert and decide

Ask how the leading plan could fail, what evidence contradicts it, who bears the loss, and what can be learned before an irreversible commitment. Distinguish reversible experiments from obligations that continue after the experiment ends. Choose a cheap test only when it can resolve a consequential uncertainty.

Make a recommendation or identify the genuine unresolved tradeoff. State the critical assumption, strongest reason, main failure mode, and specific evidence that would change the view. Probability statements need an event and resolution date; subjective confidence in advice is not that probability. Record forecasts before outcomes are known using the ledger.

Do not imply future monitoring or reminders exist because an update condition is written in a memo. Establish a scheduler or owner only when authorised and actually available. Advice to investigate, transact, hire, invest, or contact someone is not permission to execute it.

Targeted references

• Model latticework: selecting relevant lenses.
• Misjudgment: incentives and bias hypotheses.
• Checklists: domain-specific questions.
• Examples: adapted reasoning patterns.
• Arithmetic: matrix/scenario assumptions and limits.
• Portability: available tools/evidence.

The longer operating-system and memo assets remain a menu for substantial work, not a requirement to reproduce their whole structure. Keep conclusions connected to the user's actual decision and sources. No controlled decision-quality or agent-output improvement is claimed from this revision.