Differential Diagnosis

Differential Diagnosis

Enumerate the possible causes of an observed condition, order them by probability and severity, and run the tests that discriminate fastest — medicine's algorithm, generalized.

Why

Debugging-by-hunch tests hypotheses in the order they came to mind. Ordering by prior times severity times test-cost, and choosing tests for discrimination power, finds the cause faster and never silently drops the dangerous possibility.

Use when / Don't use when

• Use when: something is actively wrong, multiple causes are possible, and testing is costly — debugging, medical-style triage, churn spikes, machine faults.
• Don't use when: one cause is near-certain, or the condition is historical — use root-cause-investigation instead.

Inputs → Outputs

• Inputs: an ongoing abnormal condition in a system, organism, or organization.
• Outputs: a ranked differential, test results, the confirmed cause or narrowed set, and explicitly ruled-out dangerous causes.

Principles

• Common things are common — priors come from base rates, not memorability.
• "Can't-miss" causes get tested early even when improbable, because the cost of missing them dominates.
• The best next test is the one whose result most reshapes the ranking per unit cost.
• Diagnosis is iterative — each result re-ranks the list.

Procedure

1. Characterize the condition precisely: what's abnormal, since when, what changed around onset, what's normal — the pertinent negatives.
2. Generate the differential via abductive-inference: plausible causes plus the dangerous ones regardless of plausibility.
3. Assign priors from base rates via reference-class-forecasting; tag each cause with severity-if-missed.
4. Order the worklist by prior, promoted by severity and demoted by test cost.
5. Choose the test with the best discrimination-per-cost, preferring tests that split the list over tests that confirm the favorite.
6. Update the ranking on each result via bayesian-updating. Prune what's ruled out; add what new evidence suggests.
7. Stop when one cause explains all findings and the can't-miss causes are affirmatively excluded. Run disconfirmation-search on the final answer, and sanity-checking on the conclusion.
8. Report: cause, confidence, what was ruled out, and what re-presentation would reopen the case.

Common mistakes

• Anchoring on the first hypothesis and ordering tests to confirm it.
• Satisfaction-of-search — stopping at one cause when two are present.
• Never writing the differential down; the unlisted cause can't be found.
• Ignoring pertinent negatives.

Examples

• Intermittent API latency: GC, noisy neighbor, connection pool, DNS, retry storm.
• A patient fatigue workup.
• Sudden CAC increase: tracking bug, mix shift, competitor, seasonality.

Related

• [[root-cause-investigation]] — the post-hoc sibling for historical, resolved failures.
• [[competing-hypotheses]] — the same logic applied to beliefs instead of live systems.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: skyf0xx
• Source: skyf0xx/better-thinking
• License: MIT
• Homepage: https://zenbin.org/p/better-thinking

Install and usage instructions live in the source repository linked above.