Abductive reasoning

1. Three kinds of reasoning, three jobs

Type	Direction	Conclusion	Typical form
Deduction	general → particular	certain	All men are mortal → Socrates is mortal
Induction	particular → general	probable	Every swan seen is white → swans are probably white
Abduction	observation → cause	most likely	The grass is wet → it probably rained

Detective work is almost entirely the third kind: traces on the scene, then a reconstruction. Medical diagnosis, troubleshooting and archaeology are the same operation. Abduction is not strict logic; it is inference to the best explanation.

2. A worked example

Observation: rhythmic knocking upstairs between 23:00 and 01:00, about every 30 seconds, for ten days.

1. Enumerate: pipe expansion, a timed appliance (washer/dryer), someone moving furniture, heating circulation, a pet, construction, deliberate knocking.
2. Downgrade the implausible: deliberate knocking needs several extra assumptions (that they noticed you, cared, chose a metronome); set it aside.
3. Find the common thread: a fixed window plus a uniform interval points to something periodic, driven by a device or a routine.
4. Probability and basis: a timed appliance is the strongest candidate; the basis is the fixed timing and even spacing.
5. Design a test: log start, end and interval for three nights; compare weekends; listen at the pipe shaft.
6. Reconstruct: if it is a washer, the sound should change in stages and stop after an hour or two.

Each candidate carries "how to verify" rather than a verdict — which is what abduction should look like in practice.

3. Why it can always be overturned

The conclusion has the form "X is the best explanation", not "X is true". When new evidence arrives, the best explanation can change. That is not a defect; it is the point — you get a hypothesis that can be beaten, and therefore know what to observe next.

4. Three common errors

• Premature convergence — stopping at the first plausible idea. Enumerate first, rank second.
• Ignoring base rates — vivid explanations get overrated. See the base-rate fallacy.
• Correlation as causation — two things moving together is not proof that one drives the other. See Correlation is not causation.

5. How it fits with the razor and Bayes

Abduction generates candidates, Occam\'s razor ranks them, and Bayesian thinking updates the probabilities. Together they form the full loop — see the method.