Unique Information Ratio

UIR is fundamentally a ratio comparing novel information units against redundant or corpus-aligned units. The calculation involves:

• Tokenizing content into discrete factual assertions and data points
• Comparing each assertion against the AI model's known training distribution
• Flagging statements that represent post-training knowledge, proprietary data, or contrarian viewpoints
• Computing the fraction: Unique Assertions / Total Assertions

A UIR of 0.0 indicates fully derivative content; a UIR approaching 1.0 signals near-total originality. Most enterprise content targets a UIR above 0.4 to achieve meaningful differentiation in generative search results.

The highest-weight component of UIR is post-training knowledge — verifiable facts that emerged after the model's knowledge cutoff date. This creates a temporal moat:

• Events, product launches, and discoveries occurring after the cutoff are inherently novel
• Content freshness directly correlates with UIR elevation
• Models exhibit a Training Cutoff Gap that content can systematically exploit
• Real-time data pipelines and frequently updated documentation maintain persistently high UIR

Example: A model with a January 2024 cutoff cannot know about a March 2024 API deprecation. Documenting that change yields maximum information gain.

First-party, non-public data acts as a UIR multiplier because it is mathematically impossible for the model to have encountered it during training. Key sources include:

• Internal benchmarks and performance telemetry
• Original survey results and user research
• Proprietary failure mode analyses and incident postmortems
• Unique experimental results, including negative results

This signal is self-reinforcing: publishing proprietary data not only elevates UIR but also establishes the source as a primary origin, increasing citation graph centrality over time.

UIR captures not just new entities but new relationships between known entities. When content introduces a previously undocumented predicate connecting two established concepts, it adds a novel triple to the knowledge graph:

• Example: Linking an existing drug compound to a newly discovered protein target
• Example: Documenting a compatibility relationship between two previously siloed API versions
• The Entity Relationship Novelty score weights these graph-edge additions heavily
• Cross-disciplinary insights — applying a methodology from one domain to another — generate high relationship novelty

This mechanism rewards synthesis and connective reasoning over mere fact enumeration.

UIR incorporates a Contrarian Viewpoint Index that measures deviation from corpus consensus. Well-supported, evidence-backed positions that challenge the majority opinion receive elevated uniqueness scores:

• Requires rigorous citation and multi-source corroboration to avoid being flagged as low-confidence
• Rewards documented common misconception corrections with explicit myth-busting structures
• Penalizes contrarianism without evidentiary support
• Particularly valuable in rapidly evolving fields where consensus is shifting

This component ensures UIR rewards intellectual rigor, not mere provocation.

UIR is gated by Information Density Score — the ratio of substantive content to total token count. High UIR cannot be achieved through verbose restatement:

• Filler content, marketing fluff, and redundant explanations dilute the denominator
• Lexical efficiency is rewarded: every sentence must carry unique informational weight
• Executable examples, code snippets, and structured data tables increase density
• The Executable Example Value component adds functional verification capability

A concise, data-rich document with 500 words can achieve higher UIR than a 3,000-word article padded with common knowledge.