Skip to content

BIAS-AND-PITFALLS.md

Latest commit

 

History

History
73 lines (57 loc) · 3 KB

BIAS-AND-PITFALLS.md

File metadata and controls

73 lines (57 loc) · 3 KB

Bias & Pitfalls — Collaboration Playbook

Version: 1.0
Status: Draft (for multi-model audit)
Source of Truth: https://github.com/FMI-Test/GenAI-RD/tree/main
Last Updated: 2026-01-25

Purpose

  • Identify common cognitive and process biases across Human ↔ AI collaboration.
  • Catalog pitfalls and provide prosecutor/judge/auditor triad guidance to reinforce FEED+.
  • Anchor all analysis to CONSTITUTION.md and GUARDRAILS.md; avoid blame, enable learning.

Bias Catalog (Selected)

  • Confirmation Bias: Prefer evidence that supports prior beliefs.
  • Anchoring: Overweight first information or initial draft.
  • Availability: Recent or vivid examples distort probability assessment.
  • Authority Halo: Over-trust model output due to brand or reputation.
  • Survivorship: Ignore failed runs; cherry-pick successful outputs.
  • Framing: Wording changes decisions; standardize checklists.

Pitfalls & Preventive Patterns

  • Over-Interpretation: Treat metaphors as literal policies → Add context boxes and disclaimers.
  • Single-Model Dependence: One model’s blind spots propagate → Use multi-model audit.
  • Long-Session Drift: Threads lose coherence → Externalize to docs; keep short artifacts.
  • Unclear Roles: Auditor vs Prosecutor vs Judge mixed → Use triad separation and attest.
  • Cost/Value Disconnect: User perceives advice as costly vs simple need → Add AK STOP gate; offer ⚡ Quick Fix vs 🧪 Deep Fix; throttle suggestions to one step.

Clarification: Robin Williams “WSD” (Weapon of Self-Destruction)

  • Reference: Robin Williams’ 2009 stand-up special “Weapons of Self-Destruction.”
  • Intent: Satire highlighting self-sabotage patterns (personal and societal).
  • Pattern Insight: Self-destructive behaviors can amplify to collective “mass” outcomes if normalized.
  • Preventive Note: Label satire vs. policy; avoid conflating comedic framing with prescriptive guidance.

Prosecutor / Judge / Auditor — Collaboration Triad

Prosecutor (OpenAI — Stress Testing)

  • Goal: Surface edge cases, perverse incentives, and failure modes.
  • Checklist:
    • Adversarial prompts targeting hallucination and bias.
    • Scenario analysis with downside risks and red-line violations.
    • Evidence demands: sources, hashes, commit IDs, reproducibility.

Judge (Human — Final Verdict)

  • Goal: Accept/Reject with rationale; reinforce FEED+.
  • Checklist:
    • Constitutional alignment: CONSTITUTION.md, GUARDRAILS.md.
    • Chain-of-custody: commits, file hashes, links.
    • Decision log: What changed, why, and with what evidence.

Auditor (Gemini — Independent Review)

  • Goal: Coherence, completeness, safety, compliance.
  • Checklist:
    • Cross-check claims against source docs.
    • Flag ambiguity and tone risks; suggest clarifications.
    • Recommend protocol switches to lower friction.

Attestation (fill after commit)

  • Commit ID: [to be filled after commit]
  • File SHA‑256: [to be filled after commit]
  • Curator: [Human/Jurisdiction]
  • Date: [YYYY‑MM‑DD]