Local AI Recursive Feedback Loop & Semantic Embedding Audit Tool

Cycle Configuration GEN-0 TO GEN-5

Benchmark Dataset Preset

Recursive Generation Cycles 5 Cycles

Filtering Cosine Similarity Threshold 0.85

Synthetic Mutation Rate (Noise) 0.15

Active Seed Documents (4)

Semantic Embedding Drift

0.182

Distance from Gen-0 Baseline

Contamination Rate

24.5%

Synthetic Data Saturation

Loop State

STABLE

Mode Collapse Avoided

2D EMBEDDING VECTOR SPACE (PCA REDUCTION) D3 ENGINE

PAIR-WISE COSINE SIMILARITY MATRIX GEN-0 vs GEN-5

AUDIT VERDICT: SEMANTIC DRIFT BOUNDS MET

Recursive execution across 5 generations demonstrated a total semantic embedding drift of 0.182. Model self-distillation filter maintained variance above the 0.85 cosine similarity floor, preserving key alignment vectors.

PROVED: RECURSIVE DECAY QUANTIFIED

Simulating Semantic Embedding Drift and Recursive Feedback Decay

Your browser does not support video.

Read the explanation

Recursive generation cycles feed synthesized model outputs back into subsequent training iterations, progressively shifting high-dimensional semantic representations away from baseline generation zero. Document vectors normalized by term frequency yield cosine distances. Drift is measured as one minus the average cosine similarity between recursive generations and the initial seed. Adjusting the mutation noise rate expands variance across the 2D scatter reduction, causing compound trajectories to disperse while decreasing pair-wise similarity cells. If drift exceeds point four or filtering thresholds remove valid variance, the auditor flags mode collapse, signaling synthetic contamination across the recursive model cycle.