One-sentence brief
Metrics are useful when they expose a concrete failure pattern and dangerous when they hide context, language differences, sample size, or false-positive risk behind a single score.
AI-DRIVEN WORLD SAFETY
How to use vocabulary, entropy, opening, duplication, cluster, and behavior metrics without turning one threshold into a universal quality oracle.
ORIENTATION
Metrics are useful when they expose a concrete failure pattern and dangerous when they hide context, language differences, sample size, or false-positive risk behind a single score.
WORKING BRIEF
Use length-normalized measures for lexical diversity rather than raw type-token ratios that mechanically decline with longer text. Compare like with like across language, field, and sample length.
Track dominant first words, first phrases, full-name introductions, clause patterns, sentence length, and part-of-speech rhythms across biography and dialogue.
Measure the concentration of occupation–routine, relationship, goal, ordinary-concern, and behavioral-response combinations after semantic grouping.
A 100-character cast and a 10,000-character population need different collision expectations. Same-role cohorts may share more domain vocabulary but not identical life history or behavioral response.
Every metric should state language coverage, fields analyzed, missing data, model version, candidate recall, confidence interval, and known false-positive risks.
COMPLETE DOSSIER
Terms are defined for this site’s evidence method, not as universal legal or clinical definitions.
| Metric family | Useful for | Required caveat |
|---|---|---|
| Lexical diversity | Detecting narrow or repetitive vocabulary | Language, length, role, and genre affect baseline |
| Opening concentration | Finding formulaic greetings and biography starts | Simple sentences naturally converge |
| Cluster share | Finding dominant template families | Embedding and clustering choices shape families |
| Combination entropy | Finding tiny pools of traits or mechanics | Lore constraints can lower expected entropy |
| Review defect rate | Estimating sampled quality | Sample design and confidence interval must be shown |
Does the design preserve the exact fictional identity, ordinary life, independent goals, and ability to refuse rather than reducing the character to a role or prompt?
Pass condition: Identity fields are stable, state is separate, protected traits are not quality scores, and silent substitution is impossible.
Can every transition, validation result, accepted fingerprint, exception, and human decision be traced to a versioned record?
Pass condition: Automated checks, human review, activation authority, and production approval remain separate and explicit.
Can untrusted provider output, administrative evidence, stale revisions, or private data enter live context or binding state?
Pass condition: Only allowlisted, current, reviewed projections and bounded scene or memory packets can be used; failures degrade safely.
Can a changed source, identity revision, harmful behavior, or failed review invalidate downstream use without destroying audit history?
Pass condition: Supersession, pause, rollback, correction, and permanent retirement are defined and testable.
RESEARCH EDITION
This page follows the public method for provenance, confidence, source independence, alternative accounts, limitations, review state, and visible correction.
CONTINUE