Abductive reasoning

Abductive Reasoning

Overview

Abduction — inference to the best explanation — is the only mode of reasoning that introduces new ideas. Deduction works out consequences; induction generalizes instances; abduction generates the hypotheses themselves (Peirce, 1903).

Formal schema: The surprising fact C is observed. But if A were true, C would be a matter of course. Hence, there is reason to suspect A is true.

The "best" qualifier carries the weight: consistent-with-evidence is not enough — the hypothesis must beat rivals on coverage, simplicity, prior probability, and predictive power (Harman 1965; Lipton 2004).

Composes with bayesian-reasoning (Bayes scores hypotheses abduction generates), occams-razor (simplicity criterion), critical-thinking (competing-hypotheses analysis), and debugging-and-error-recovery (technical debugging is iterated abduction).

When to Use

• A surprising observation needs explanation (bug, symptom, outage, customer behavior, financial anomaly)
• A diagnostic decision must be made under uncertainty (medical, technical, investigative)
• Someone is treating "consistent with X" as proof of X
• A team has converged on one explanation without enumerating alternatives
• An investigation has stalled at the first plausible-sounding hypothesis

Not when: the problem is deductive (math, formal logic); evidence is sufficient for direct measurement; you have no domain knowledge to generate candidate hypotheses.

Coaching Novices (Adaptive Front Door)

• Engine mode: user has a concrete case → run The Process directly.
• Coach mode: user is unfamiliar or has no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. One-line what-it-is: instead of "this looks like X, so it's X," abduction enumerates every hypothesis that would explain the observation and picks the one that best accounts for all the evidence — coverage, simplicity, prior plausibility, predictive power.
2. Check fit: deductive / direct-observation problems → not this lens.
3. Elicit the surprising observation — what happened that didn't match expectations?

[WAIT — do not advance until user responds]

4. Run The Process one step at a time with their input — generate alternatives, score each on the four criteria.

[WAIT — do not advance until user responds]

5. Close: name the chosen hypothesis, the closest rival, and the test that would distinguish them.

[WAIT — do not advance until user responds]

The Process

S1 — State the surprising observation: what was expected vs what happened. S2 — Generate candidates: H1…H5 + H_unknown (the explanation not yet considered). Force ≥3. Techniques: domain expertise, inversion, failure modes, adversarial generation, base rates. S3 — Score each on four criteria: Coverage (explains all evidence?), Simplicity (fewest auxiliary assumptions?), Prior (base-rate plausibility?), Predictive power (new testable predictions?). S4 — Select the best: highest combined score, not just "most consistent." State why it beats the runner-up. S5 — Name the discriminating test: what evidence would confirm it, what would refute it, what would distinguish it from the runner-up. S6 — Commit provisionally: treat as current best hypothesis. Specify an update trigger; if the test fails, return to S2.

Output: Abductive Inference

# Abductive Inference: <surprising observation>
Observation: <what> / Why surprising: <expected vs actual>
Candidates: H1 | H2 | H3 | H_unknown
| H | Coverage | Simplicity | Prior | Predictive power |
|---|---|---|---|---|
Best explanation: <chosen> — why over runner-up: <…>
Discriminating test: predicts <…> / refuted by <…> / distinguishes from runner-up by <…>
Action: provisional decision <…> / update trigger <…>

→ Method in Action: Peirce, the Wine Cellar, and the Watch — 1879 · The Discovery of Neptune — 1846

Pack: Abduction in Practice

Domain	Surprising observation	Common abductive failure
Medical diagnosis	Unusual symptom cluster	Settling on first matching diagnosis; ignoring prior
Software debugging	Test fails unexpectedly	Blaming the change; ignoring environment/build causes
Business performance	Revenue drop in a segment	Blaming sales without checking market/product/measurement

Applying It Well

• Generate broadly before committing narrowly — most failures are failures to enumerate
• "Consistent with X" is the weakest criterion; always compare across all four criteria
• State the conclusion as provisional; name the update trigger — → Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Fake move	Reality
[D] "The first explanation that fits is correct"	The first explanation rarely is. Abduction requires enumeration.
[D] "This is consistent with X, so X"	Consistency is necessary, not sufficient. Compare against rivals.
[D] "I can't think of another explanation"	Generation step is incomplete. Use inversion, base rates, adversarial generation.
[D] Picking a hypothesis because it's vivid or interesting	Storytelling fallacy. Vividness is not evidence. Apply the four criteria.
[D] Treating abductive conclusion as proven	Abduction is provisional. Always state the discriminating test.
[D] Skipping the H_unknown placeholder	Commits to the best hypothesis you happened to think of — absence of imagination ≠ absence of possibility.
[D] Selecting based on coverage alone	Coverage is one of four criteria. High-coverage + wildly improbable prior = still wrong.
[D] "Abduction is the same as guessing"	Abduction is structured: enumerate, score, select. Guessing skips the scoring.
→ Add [O] entries here after each real use — paste the actual failure pattern	What went wrong and why

Red Flags

• Only one hypothesis was considered
• Selected hypothesis was the first that "matched"
• No comparison across multiple criteria
• No discriminating test stated
• No H_unknown placeholder
• Confidence framed as proof rather than best-current-explanation
• Explanation requires auxiliary coincidences without acknowledging them

Verification

• At least 3-5 candidate hypotheses generated, plus H_unknown
• Each scored on coverage, simplicity, prior, predictive power
• Selected hypothesis compared to runner-up explicitly
• Discriminating test identified
• Conclusion stated as provisional, not proven
• Update trigger specified

---

Part of deciqAI Knowledge Skills — 233 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/s/abductive-reasoning · Built by deciqAI · github.com/deciqAI · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/abductive-reasoning.json