≈1.3k tokens on demand. Your agent reads this file only when SKILL.md points to it.
Purpose: Use this reference to detect and reduce cognitive bias in strategic work. It preserves the top bias patterns, observed rates, and practical debiasing mechanisms Helm should apply.
Contents
CB-01..CB-10- Debiasing methods
- Phase risk map
- Helm integration
Cognitive Biases in Strategic Decision-Making
Top Biases
| ID |
Bias |
Observed rate |
Strategic impact |
Debiasing move |
CB-01 |
Confirmation bias |
78.2% |
Evidence is filtered to support the preferred thesis |
Use neutral fact-building and counterevidence review |
CB-02 |
Overconfidence |
71.0% |
Complexity and downside are underestimated |
Run a pre-mortem |
CB-03 |
Anchoring |
59.6% |
First framing dominates decision quality |
Evaluate multiple anchors in parallel |
CB-04 |
Loss aversion |
56.0% |
Necessary exits or pivots are delayed |
Make opportunity cost explicit |
CB-05 |
Action bias |
— |
Teams act before thinking deeply enough |
Force a deliberate evaluation window |
CB-06 |
Planning fallacy |
— |
Time, cost, and risk are underestimated |
Compare with external reference cases |
CB-07 |
Groupthink |
— |
Harmony blocks critique and diversity |
Require dissent and structured review |
CB-08 |
Dunning-Kruger effect |
— |
Teams overestimate internal capability |
Use benchmarking and 360 feedback |
CB-09 |
Framing effect |
— |
Choice changes with wording |
Reframe the same case multiple ways |
CB-10 |
Sunflower bias |
— |
Teams align with leader preference instead of evidence |
Anonymous input, leader speaks last |
Debiasing Toolkit
Structural Interventions
| Tool |
Use |
Best for |
| Red Team / Devil’s Advocate |
Build the strongest argument against the preferred path |
confirmation bias, groupthink |
| Pre-mortem |
Assume failure first, then work backward |
overconfidence, planning fallacy |
| External perspective |
Bring in benchmarks or independent review |
confirmation bias, anchoring |
| Anonymous voting / Delphi |
Collect judgments before social influence kicks in |
groupthink, sunflower bias |
AI- and Data-Assisted Debiasing
| Signal |
Use |
| Bias pattern detection |
Detect language patterns that imply overconfidence or cherry-picking |
| Scenario diversity scoring |
Flag scenario sets that are too homogeneous |
| Accuracy tracking |
Feed FORESIGHT to see where bias keeps recurring |
| Tool-assisted debiasing |
Research suggests up to 16% improvement in strategic outcomes |
Phase Risk Map
| Phase |
Highest-risk biases |
Mandatory check |
| Environment analysis |
confirmation bias, availability |
search for disconfirming evidence |
| Goal setting |
overconfidence, planning fallacy |
compare against outside benchmarks |
| Strategy design |
anchoring, groupthink |
run Red Team review |
| Decision-making |
loss aversion, framing |
restate options from multiple frames |
| Execution planning |
planning fallacy, action bias |
compare with similar prior cases |
| Monitoring |
confirmation bias, sunk-cost logic |
predefine kill criteria |
Helm Integration
- Apply a bias check in every
SCAN, MODEL, SIMULATE, and ROADMAP pass.
- Standardize pre-mortems in
SIMULATE.
- Feed detected patterns into
FORESIGHT.
- Score scenario diversity before finalizing a strategy package.
- Include a compact bias risk map in strategic review output when uncertainty is high.
2025-2026 Notes
- Daniel Kahneman (1934-2024) — Thinking, Fast and Slow (2011) author and 2002 Nobel laureate — died 2024-03-27, age 90 (Princeton 2024-03-28, NPR 2024-03-27). The System 1 / System 2 framing remains the dominant pop-cognitive model; Richard Thaler (Nobel 2017, Nudge) and Cass Sunstein continue active publication. There is no single anointed successor — treat the field as plural and beware "argument from authority" framings of Kahneman quotes.
- AI as bias amplifier and bias-detector — both — Helm should:
- Use LLMs to flag confirmation-laden phrasing, monolithic scenario sets, and one-sided benefit framing before sign-off.
- Avoid using LLMs as the arbiter of which scenario is "most likely" — RLHF tends to produce overconfident, conventional-wisdom-skewed completions.
- When facilitating Red Team / Devil's Advocate, instruct the model to defend the minority view with citations and reject "balanced both-sides" output as a sunflower-bias artifact.
- WEIRD critique extension (2025-): training data for both human-judgment baselines and LLM priors remains heavily Western / English / late-2010s-internet — apply geographic and temporal debiasing when the simulation crosses borders or addresses pre-2020 historical analogs.
