deciqAI Critical Thinking
Overview
Critical thinking evaluates a conclusion by explicitly testing it against alternative interpretations — the conclusion that survives the most disconfirmation attempts is the one to act on, not the one with the most confirming examples. Operationalized by Richards J. Heuer Jr. (CIA, 1999) as Analysis of Competing Hypotheses (ACH): a matrix marking each evidence-hypothesis pair as Consistent / Inconsistent / N/A. The hypothesis with the fewest Inconsistencies wins.
Compose: first-principles (surface assumptions) · probabilistic-thinking (calibrate evidence) · inversion (generate alternatives). Critical thinking = evaluation; first-principles = construction; occams = selection; second-order = projection.
When to Use
- High-stakes claim (yours or someone else's) that needs evaluation
- Multiple interpretations of the same evidence are possible
- Motivated reasoning detected — conclusion shaped by what you want to be true
- Team converging on a hypothesis too fast without seriously considering alternatives
- Evaluating an AI capability or AI-hype claim (e.g. "this benchmark proves the model reasons") where the same evidence fits several explanations (contamination, cherry-picking, teaching-to-the-eval)
When NOT to use: trivial/already-validated claim; only one hypothesis imaginable (use first-principles first); time pressure makes 90-min ACH impossible; values question not a facts question.
Coaching Novices (Adaptive Front Door)
- Engine mode: user has a concrete claim + evidence → 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.
- One-line what-it-is: critical thinking is testing your conclusion against alternative explanations, looking for the one with the fewest unresolved contradictions — not the most confirming examples.
- Check fit against When to Use / When NOT to use. Redirect if trivial or values-only.
- Elicit the specific claim under evaluation — "I want to evaluate my thinking" is too vague; need a specific claim. > [WAIT — do not advance until user responds]
- Walk through ACH steps one at a time with their input: list hypotheses → list evidence → score matrix → identify survivor. > [WAIT — do not advance until user responds]
- Close by naming the most-diagnostic evidence they discovered — the piece that most differentiates hypotheses. > [WAIT — do not advance until user responds]
The Process
Run Analysis of Competing Hypotheses (ACH) — Heuer's 8-step procedure.
- Identify hypotheses. Brainstorm broadly — obvious, contrarian, and 2-3 in between. Generate at least 3 alternatives.
- List significant evidence for and against each hypothesis — include absence of expected evidence.
- Prepare the matrix: hypotheses across top, evidence down side. Mark each cell C / I / N/A. Load-bearing step.
- Refine the matrix. Remove evidence consistent with all hypotheses (no discriminating value). Focus where hypotheses disagree.
- Draw tentative conclusions. Count Inconsistent marks — the hypothesis with the fewest I marks is the survivor.
- Sensitivity analysis. If the most-important evidence were wrong, would the conclusion change? Identify linchpin assumptions.
- Report conclusions — alternatives considered, why rejected, and linchpin assumptions explicitly stated.
- Set milestones for future observation that may indicate events are taking a different course.
Output: the ACH Matrix
Hypotheses: H1 / H2 / H3 (≥3)
Evidence: E1, E2, E3 … (include absence of expected evidence)
Matrix: mark each cell C / I / N/A
Inconsistencies count: H1=2, H2=3, H3=0 → H3 survives
Linchpin assumptions: "If E2 were wrong, conclusion flips to H1"
Alternatives rejected: H1 because E2+E5; H2 because E1+E4
Milestones: "If X observed by <date>, re-run matrix"
→ Method in Action: ACH at the CIA and the Cuban Missile Crisis Pattern (1962 → 1999) · Piltdown Man Forgery Debunked (1912 → 1953) → 2026 lens: "This Benchmark Proves the Model Reasons" — an ACH on AI Capability Claims (2024–2026)
Analysis Packs
Domain packs (canonical hypotheses, diagnostic evidence, typical bias, matrix conventions):
- Intel: state intentions/capabilities; HUMINT/SIGINT/IMINT; bias = mirror-imaging.
- Medical: differentials; history + labs + imaging; bias = premature closure.
- Business/investment: market thesis alternatives; data + qualitative signals; bias = motivated reasoning.
- Engineering: root-cause candidates; logs + repro steps; bias = fixing the visible symptom layer.
Applying It Well
- Alternatives before evidence — listing evidence first lets your initial hypothesis shape selection.
- Fewest Inconsistencies beats most Consistencies — Heuer's central insight.
- Diagnostic evidence > confirmatory — evidence consistent with all hypotheses doesn't discriminate.
- Name linchpin assumptions — without naming what would change your mind, the analysis is not falsifiable.
- Re-run as evidence accumulates — Cuba 1962 failure was partly about never revisiting NIE 85-3-62.
→ Primary sources: references/sources.md
Common Rationalizations
[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.
| Fake move | Reality |
|---|---|
| [D] Only one hypothesis considered | The most common failure. If "there aren't any plausible alternatives," generate them via first-principles before continuing. |
| [D] Counting Consistent marks instead of Inconsistent marks | Confirmatory evidence often supports multiple hypotheses; what differentiates is the disconfirming evidence. |
| [D] Skipping the matrix because "I can do this in my head" | The matrix exists because unaided reasoning systematically anchors on the first hypothesis. |
| [D] Discarding evidence that contradicts the favored hypothesis | Mark it Inconsistent. Discarding it means the matrix is rigged. |
| [D] Confusing absence of evidence with evidence of absence | If H1 predicts X and you don't see X, that's diagnostic — mark Inconsistent for H1. |
| [D] No linchpin assumptions stated | If you cannot name what would change your conclusion, it is not falsifiable. |
| [D] Confirmation bias laundered as "intuition" | "My intuition says H1" + writing up analysis to confirm H1 is the Sherman Kent 1962 failure mode. |
| [D] Motivated reasoning about the cost of being wrong | Hypotheses implying painful actions are systematically under-weighted. Recognize and over-correct. |
| [D] One-shot analysis, never revisited | New evidence requires re-running. The Cuba failure was partly about not revisiting NIE 85-3-62. |
| → Add [O] entries here after each real use — paste the actual failure pattern | What went wrong and why |
Red Flags
- Only one hypothesis is being considered
- "Supporting evidence" accumulated without checking against alternatives
- Conclusion reached before the analysis (motivated reasoning)
- No matrix or equivalent structured comparison exists
- No statement of linchpin assumptions / what would change the conclusion
- Disconfirming evidence explained away ("yes, but...") rather than entered into the matrix
- Analysis done once and never revisited as new evidence arrived
Verification
- At least 3 hypotheses are explicitly listed
- Evidence is enumerated independently of any hypothesis preference
- Each evidence-hypothesis pair is marked C / I / N/A
- The matrix focuses on Inconsistencies, not Consistencies
- Diagnostic evidence (where hypotheses disagree) is highlighted
- Surviving hypothesis is the one with fewest Inconsistencies
- Linchpin assumptions are explicitly named
- Milestones for re-evaluation are set
Part of deciqAI Knowledge Skills — 237 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/critical-thinking · Built by deciqAI · github.com/deciqAI · Contributions welcome.
Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/critical-thinking.json