bias-candidates

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:

1. supporting episodes (with dates and platforms)
2. counterexamples found on purpose
3. alternative explanations that fit the same data
4. confidence: HIGH / MEDIUM / LOW / SPECULATIVE
5. practical consequence if true
6. cheapest interruption strategy
7. 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.