references/cognitive_bias_canon.md
8.4 KB查看排版效果| 1 | # Cognitive Bias Canon — 5 Biases + Recognition Cues |
| 2 | |
| 3 | This reference answers exactly one decision: **which 5 cognitive biases does the reflect skill check for, and how does each manifest in conversation patterns?** |
| 4 | |
| 5 | ## The Core Frame |
| 6 | |
| 7 | 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. |
| 8 | |
| 9 | ## The 5 Biases |
| 10 | |
| 11 | | Bias | Definition | Conversation signal | Corrective | |
| 12 | |---|---|---|---| |
| 13 | | **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 | |
| 14 | | **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 | |
| 15 | | **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 | |
| 16 | | **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 | |
| 17 | | **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 | |
| 18 | |
| 19 | ## 1. Confirmation Bias |
| 20 | |
| 21 | Wason (1960) demonstrated that people systematically seek confirming evidence over disconfirming. In conversation, this manifests as: |
| 22 | |
| 23 | - **Selective citation:** "X supports our hypothesis" without checking for counter-cases |
| 24 | - **Asymmetric scrutiny:** confirming evidence accepted; disconfirming evidence questioned |
| 25 | - **Strawmanning alternatives:** weak versions of opposing positions cited |
| 26 | |
| 27 | ### Recognition in conversation |
| 28 | |
| 29 | Look for: 3+ supporting examples cited with no counter-examples; phrases like "everything we've found supports..."; competing hypotheses absent or only weakly framed. |
| 30 | |
| 31 | ### Corrective move |
| 32 | |
| 33 | 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. |
| 34 | |
| 35 | ## 2. Sunk Cost Fallacy |
| 36 | |
| 37 | Arkes & Blumer (1985) showed people irrationally continue based on prior investment. In conversation: |
| 38 | |
| 39 | - **"We're far enough in to..."** signals sunk-cost reasoning |
| 40 | - **"After all that work..."** — past effort treated as locked-in value |
| 41 | - **Switching cost weighted higher than continuation cost** without specific calculation |
| 42 | |
| 43 | ### Recognition in conversation |
| 44 | |
| 45 | 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." |
| 46 | |
| 47 | ### Corrective move |
| 48 | |
| 49 | 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. |
| 50 | |
| 51 | ## 3. Anchoring |
| 52 | |
| 53 | Tversky & Kahneman (1974) demonstrated that initial estimates persist even when irrelevant. In conversation: |
| 54 | |
| 55 | - **First option becomes the default frame** even when better alternatives emerge |
| 56 | - **"Compared to X..."** when X was the first option — alternatives evaluated relative to anchor, not absolutely |
| 57 | - **Range-bound thinking** around the anchor's neighborhood |
| 58 | |
| 59 | ### Recognition in conversation |
| 60 | |
| 61 | Look for: multiple options surfaced but discussion keeps circling back to the first; alternatives framed as "modifications of X" rather than fundamentally different approaches. |
| 62 | |
| 63 | ### Corrective move |
| 64 | |
| 65 | "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. |
| 66 | |
| 67 | ## 4. Complexity Bias |
| 68 | |
| 69 | The opposite of Occam's razor — adding layers because they sound rigorous, not because each is justified. In conversation: |
| 70 | |
| 71 | - **Each layer plausible in isolation** — but cumulative complexity exceeds problem complexity |
| 72 | - **Safeguards / wrappers / fallbacks** added speculatively without specific failure mode |
| 73 | - **"What about..." additions** without "would dropping this break anything?" check |
| 74 | |
| 75 | ### Recognition in conversation |
| 76 | |
| 77 | Look for: a feature/layer/check added without naming the specific failure it prevents; cumulative architecture growing turn-over-turn without consolidation. |
| 78 | |
| 79 | ### Corrective move |
| 80 | |
| 81 | 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. |
| 82 | |
| 83 | ## 5. Recency Bias |
| 84 | |
| 85 | The last N turns dominate working memory; turns 1-5 fade. In conversation: |
| 86 | |
| 87 | - **Original goal forgotten** — work moved on, original constraint dropped |
| 88 | - **Recent micro-decisions cited** as if they were core principles |
| 89 | - **Strategic context** (set early) supplanted by tactical context (set late) |
| 90 | |
| 91 | ### Recognition in conversation |
| 92 | |
| 93 | 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. |
| 94 | |
| 95 | ### Corrective move |
| 96 | |
| 97 | 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. |
| 98 | |
| 99 | ## When Multiple Biases Are Detected |
| 100 | |
| 101 | In long conversations, 2-3 biases often surface together. Pattern: |
| 102 | |
| 103 | - **Confirmation + sunk cost** = "we're invested AND it's working" (resist pivoting even when alternatives are stronger) |
| 104 | - **Anchoring + complexity** = "the first idea, with N safeguards" (over-engineered version of first option) |
| 105 | - **Recency + complexity** = recent additions become core; original simple goal forgotten |
| 106 | |
| 107 | Surface each bias separately. Don't conflate. Each has a different corrective. |
| 108 | |
| 109 | ## Operational Checklist (Per Reflection) |
| 110 | |
| 111 | For each of the 5 biases: |
| 112 | |
| 113 | - [ ] Scan conversation for signal patterns |
| 114 | - [ ] If detected: name the bias, cite specific conversation evidence (with turn numbers if possible), suggest the corrective |
| 115 | - [ ] If not detected: state explicitly that you checked and didn't find it (so user knows you didn't skip the check) |
| 116 | |
| 117 | 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. |
| 118 | |
| 119 | ## Citations (7 sources) |
| 120 | |
| 121 | 1. **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. |
| 122 | |
| 123 | 2. **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. |
| 124 | |
| 125 | 3. **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. |
| 126 | |
| 127 | 4. **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. |
| 128 | |
| 129 | 5. **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. |
| 130 | |
| 131 | 6. **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. |
| 132 | |
| 133 | 7. **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. |
文件内容是平台检查时读出的文本，只供查看。平台不会运行包里的任何脚本或程序。
