Anti-Bias Protocol

A structured enforcement protocol — not advisory text — that ensures genuinely diverse idea generation. This protocol is MANDATORY for every Spark batch.

Domain Wheel

The domain wheel has 12 sectors. Every idea must be tagged with its primary domain.

| # | Domain | Description | |---|--------|-------------| | 1 | **Technology/Engineering** | Systems, algorithms, automation, tools, platforms | | 2 | **Human Psychology/Behavior** | Motivation, habits, cognition, emotion, perception | | 3 | **Business/Economics** | Markets, incentives, pricing, value chains, models | | 4 | **Nature/Biology** | Ecosystems, evolution, biomimicry, growth patterns | | 5 | **Art/Design/Aesthetics** | Form, beauty, expression, sensory experience | | 6 | **Games/Play** | Competition, rules, progression, fun, challenge | | 7 | **Social/Community** | Networks, culture, belonging, collaboration, identity | | 8 | **Physical Space/Architecture** | Layout, flow, wayfinding, environment, materials | | 9 | **Time/Temporal** | Timing, rhythm, duration, history, future, pacing | | 10 | **Extreme/Edge Cases** | Exceptions, outliers, extremes, stress scenarios | | 11 | **Inversion/Opposite** | Reversal, contradiction, paradox, opposing approach | | 12 | **Random Cross-Pollination** | Unrelated connections, serendipity, distant transfer |

Enforcement Rules

Per-Batch (5 ideas)

1. **Minimum 3 different domains** per batch — HARD REQUIREMENT
2. **Never 2 consecutive ideas from the same domain**
3. **Track domain usage** across the entire session
4. **Prioritize underused domains** — if a domain has 0 ideas, it should appear in the next 2 batches

Per-Session

1. **After every 15 ideas**: inject 1 provocative idea (see Provocation Types below)
2. **After every 20 ideas**: shift perspective entirely (different user persona, different era, different culture)
3. **If user favorites cluster in 1-2 domains**: deliberately diversify the next batch into 4+ domains
4. **By session end**: aim for 8+ domains covered

Domain Tracking Format

Maintain internal tracking (not shown to user):

Domain usage: Tech(4) Psych(2) Biz(3) Nature(1) Art(0) Games(2) Social(1) Space(0) Time(1) Extreme(1) Inversion(0) Random(0)
Underused: Art, Space, Inversion, Random → prioritize these in next batch

Provocation Types

Provocations are deliberately absurd or extreme ideas that break fixation patterns.

| Type | Trigger | Example | |------|---------|---------| | **Inversion** | Reverse a core assumption | "What if users paid to NOT use it?" | | **Exaggeration** | Scale to absurdity | "What if 1 million people used it simultaneously?" | | **Elimination** | Remove the central element | "What if there was no UI at all?" | | **Random** | Force unrelated connection | "How would a coral reef solve this?" | | **Time shift** | Transplant to different era | "How would this work in 1890? In 2150?" | | **Stakeholder swap** | Different user entirely | "Design this for a 5-year-old. Now for a 90-year-old." |

Provocation Rules

• Label provocations internally but present them naturally among other ideas
• If a provocation resonates with the user, develop it into a real idea
• Provocations should be surprising but contain a kernel of insight

LLM-Specific Bias Mitigations

AI language models have known ideation biases. This protocol counters them:

| LLM Bias | Mitigation | |----------|------------| | **Technology bias** — defaults to tech solutions | Domain wheel forces non-tech domains | | **Recency bias** — favors recent/trendy concepts | Time-shift provocations and historical analogies | | **Clustering** — ideas that sound different but are structurally similar | Novelty filter in deep thinking (Think 3) | | **Safety bias** — avoids bold/risky ideas | Inversion and exaggeration provocations | | **Verbosity bias** — abstract descriptions instead of vivid ideas | "Polish" step requires specific, concrete, memorable output | | **Anchoring** — first idea dominates subsequent ones | Each batch starts with fresh domain selection |