Persona Anchors & Invariance
Deep dive into the 6-point latent anchor protocol, mathematical verification gates, and drift prevention.
In generative persona engineering, preserving identity across hundreds of iterations requires mathematical boundaries. Kanna enforces six immutable anchors.
The Six Anchor Coordinates
| Anchor Zone | Strict Feature Locked | Forbidden Alterations |
| :--- | :--- | :--- |
| 01. Face | East-Asian anime heroine, heart-shaped jawline, calm subtle smile | Chibi distortion, caricature, western facial geometry |
| 02. Hair | Chin-length messy bob, navy #1A1E2A with electric blue #3D6BFF highlights | Brown drift, blonde streaks, long hair |
| 03. Irises | Amber-gold #E8A030 with dual catchlights | Red pupils (red reserved for collar only), green irises |
| 04. Earring | Right-ear sterling chevron > dangle on chain | Plain hoops, gold rings, multiple piercings |
| 05. Attire | Oversized cream ribbed cardigan slipped off shoulder, sky-blue ruffled camisole | T-shirts, heavy jackets, formal suits |
| 06. Familiar | Black cat with amber eyes, red collar, silver circular - tag | Dogs, missing collar, missing tag |
Verification Pipeline
Before any visual frame is accepted by the daemon, an automated feature verification pass checks the latent embedding:
fn verify_anchors(latent: &LatentTensor) -> Result<AnchorScore, DriftError> {
let score = compute_cosine_similarity(latent, &KANNA_CANONICAL_ANCHORS);
if score < 0.985 {
return Err(DriftError::AnchorRegressionDetected);
}
Ok(score)
}
If any anchor drops below 98.5% confidence, the frame is rejected instantly and re-seeded.