| name | bias-candidates |
| description | Use this to test decision patterns against repeated behavioral evidence - without ever labeling from one anecdote. |
Bias Candidates
The Bar
Never label a bias from one vivid conversation.
A candidate only enters the report after repeated behavioral evidence
across independent episodes. Required language:
"POSSIBLE RECURRING PATTERN CONSISTENT WITH anchoring"
Not "you are anchored." This is intervention material, not diagnosis.
The Candidate List
Test against observed episodes:
- anchoring on first numbers, plans, or explanations
- confirmation seeking over discriminating tests
- premature closure - stopping search at first workable answer
- sunk-cost continuation past the evidence
- availability effects - recent failures overweighted
- automation bias - accepting plausible AI output without checks
- action bias - doing instead of deciding under uncertainty
- escalation of commitment as deadline pressure rises
- over-reliance on a confident-sounding AI explanation
The Candidate Card
For each survivor include all seven fields:
- supporting episodes (with dates and platforms)
- counterexamples found on purpose
- alternative explanations that fit the same data
- confidence: HIGH / MEDIUM / LOW / SPECULATIVE
- practical consequence if true
- cheapest interruption strategy
- what prospective data would confirm or kill it
Ground terminology in the literature in SOURCES.md - Heuer for hypothesis
competition, Klayman and Ha for positive-test strategy, Parasuraman and
Manzey for automation bias, Croskerry for premature closure.
When It Backfires
- Bias-hunting as flattery-in-reverse. Two candidates confirmed honestly
beat nine asserted carelessly.
- Ignoring adaptive value. Some "biases" are correct heuristics in this
subject's actual environment. Kahneman and Klein's expertise conditions
apply: valid cues plus fast feedback.
One-Line Memory
Patterns repeat or they don't exist.
