references/cognitive_bias_canon.md
Cognitive Bias Canon — 5 Biases + Recognition Cues
This reference answers exactly one decision: which 5 cognitive biases does the reflect skill check for, and how does each manifest in conversation patterns?
The Core Frame
A reflection that doesn't check for cognitive bias is just a summary. The 5 biases below are the most operationally relevant for in-conversation reflection — each is detectable from specific conversational signals + correctable with a specific next move.
The 5 Biases
| Bias | Definition | Conversation signal | Corrective | |---|---|---|---| | Confirmation | Seeking evidence that supports the working hypothesis; ignoring counter-evidence | Cited evidence one-sided; counter-evidence dismissed or absent | Run a disconfirming-evidence pass | | Sunk cost | Continuing because of past investment, not future expected value | "We've already invested X" / "too far along to change" | Re-frame: ignore past investment, compute future value from current state | | Anchoring | Stuck on first option mentioned; alternatives compared against anchor rather than evaluated independently | Multiple options discussed but always against the first one | Re-evaluate each option on its own merits, blind to ordering | | Complexity bias | Adding features, steps, safeguards without specific justification for each | Each layer added is plausible but cumulatively bloated | Force "why this specifically, not without it?" per layer | | Recency bias | Over-weighting last few turns; older important context being ignored | Recent details cited; original goal forgotten | Re-read from turn 1, not just the tail |
1. Confirmation Bias
Wason (1960) demonstrated that people systematically seek confirming evidence over disconfirming. In conversation, this manifests as:
- Selective citation: "X supports our hypothesis" without checking for counter-cases
- Asymmetric scrutiny: confirming evidence accepted; disconfirming evidence questioned
- Strawmanning alternatives: weak versions of opposing positions cited
Recognition in conversation
Look for: 3+ supporting examples cited with no counter-examples; phrases like "everything we've found supports..."; competing hypotheses absent or only weakly framed.
Corrective move
Ask: "What would falsify this? What's the strongest counter-case we haven't engaged with?" Run a disconfirming-evidence search. The dossier skill's ≥30% disconfirming rule is this discipline operationalized.
2. Sunk Cost Fallacy
Arkes & Blumer (1985) showed people irrationally continue based on prior investment. In conversation:
- "We're far enough in to..." signals sunk-cost reasoning
- "After all that work..." — past effort treated as locked-in value
- Switching cost weighted higher than continuation cost without specific calculation
Recognition in conversation
Look for: explicit references to past investment without future-value calculation; resistance to pivoting that's framed by "we've already X" rather than "the alternative isn't better."
Corrective move
Force this reframe: "If we were starting fresh today, with current information, would we still choose this path?" If no → pivot. Past investment is irrelevant to future decisions.
3. Anchoring
Tversky & Kahneman (1974) demonstrated that initial estimates persist even when irrelevant. In conversation:
- First option becomes the default frame even when better alternatives emerge
- "Compared to X..." when X was the first option — alternatives evaluated relative to anchor, not absolutely
- Range-bound thinking around the anchor's neighborhood
Recognition in conversation
Look for: multiple options surfaced but discussion keeps circling back to the first; alternatives framed as "modifications of X" rather than fundamentally different approaches.
Corrective move
"Forget the first option. If you saw these alternatives fresh, which would you pick on its merits?" The blind-comparison technique decouples evaluation from anchoring.
4. Complexity Bias
The opposite of Occam's razor — adding layers because they sound rigorous, not because each is justified. In conversation:
- Each layer plausible in isolation — but cumulative complexity exceeds problem complexity
- Safeguards / wrappers / fallbacks added speculatively without specific failure mode
- "What about..." additions without "would dropping this break anything?" check
Recognition in conversation
Look for: a feature/layer/check added without naming the specific failure it prevents; cumulative architecture growing turn-over-turn without consolidation.
Corrective move
Per layer: "What specific failure does this prevent? What goes wrong if we drop it?" If answer is vague, drop it. The Karpathy-coder discipline in this repo (engineering/karpathy-coder/) is this corrective formalized.
5. Recency Bias
The last N turns dominate working memory; turns 1-5 fade. In conversation:
- Original goal forgotten — work moved on, original constraint dropped
- Recent micro-decisions cited as if they were core principles
- Strategic context (set early) supplanted by tactical context (set late)
Recognition in conversation
Look for: framing that references "what we've been working on" without referencing "what we were trying to accomplish"; absence of the original goal statement when justifying current direction.
Corrective move
Re-read from turn 1. State the original goal explicitly. Compare current direction to original goal. This is the discipline that distinguishes the reflect skill from a local-context summary.
When Multiple Biases Are Detected
In long conversations, 2-3 biases often surface together. Pattern:
- Confirmation + sunk cost = "we're invested AND it's working" (resist pivoting even when alternatives are stronger)
- Anchoring + complexity = "the first idea, with N safeguards" (over-engineered version of first option)
- Recency + complexity = recent additions become core; original simple goal forgotten
Surface each bias separately. Don't conflate. Each has a different corrective.
Operational Checklist (Per Reflection)
For each of the 5 biases:
- [ ] Scan conversation for signal patterns
- [ ] If detected: name the bias, cite specific conversation evidence (with turn numbers if possible), suggest the corrective
- [ ] If not detected: state explicitly that you checked and didn't find it (so user knows you didn't skip the check)
The 5-bias check is the most under-performed step in casual reflection. Doing it carefully is what separates real reflection from rationalizing the current path.
Citations (7 sources)
-
Tversky, A. & Kahneman, D., "Judgment under Uncertainty: Heuristics and Biases" — Science 185(4157), 1974, pp. 1124-1131. Foundational paper. Source for anchoring + several other biases the skill checks. The 50-year-old methodology still defines how we recognize these in real reasoning.
-
Kahneman, D., Thinking, Fast and Slow (FSG, 2011). Synthesis of decades of bias research. Source for the System-1-vs-System-2 framing that justifies reflection as a deliberate System-2 intervention against System-1 bias.
-
Wason, P. C., "On the failure to eliminate hypotheses in a conceptual task" — Quarterly Journal of Experimental Psychology 12(3), 1960. Foundational confirmation bias paper. The "2-4-6 task" showed people systematically test confirming hypotheses.
-
Arkes, H. R. & Blumer, C., "The psychology of sunk cost" — Organizational Behavior and Human Decision Processes 35(1), 1985. Empirical paper on sunk cost. Source for the "ignore past investment in future decisions" corrective.
-
Russo, J. E. & Schoemaker, P. J. H., Decision Traps (Doubleday, 1989). Practitioner-oriented synthesis of decision biases. Source for the "blind-comparison" technique that counters anchoring.
-
Tetlock, P., Superforecasting (Crown, 2015). Empirical evidence that "active open-mindedness" (Tetlock's term) is the #1 trait of accurate forecasters. The reflect skill's bias-check discipline is an operationalization of this trait.
-
Karpathy, A., "Software 2.0" + various blog posts on engineering discipline. Source for the complexity-bias corrective ("what specific failure does each layer prevent?"). The Karpathy-coder skill in this repo formalizes this.