Exploratory Embedding Analysis

This is a compact exploratory analysis of serialized AX-CPT representations. It should not be treated as evidence about latent model states or mechanistic representations.

Inputs

outputs/condition_level_representations.jsonl: 8 rows.
outputs/sliding_window_representations.jsonl: 4800 rows.

Trial-level representations are not embedded in this first pass.

Embedding Model

Model/library: local_hashing_token_ngram_v1 implemented locally in scripts/run_embedding_analysis.py.
Dependency: numpy for vector math, cosine similarity, and PCA.
Text processing: lowercase alphanumeric tokenization with regex [a-z0-9_]+.
Features: token unigrams and bigrams.
Vectorization: deterministic signed feature hashing with CRC32 into 256 dimensions.
Normalization: L2 normalization per row.

These are actual text-derived embedding vectors for the serialized representations. They are not neural embeddings, latent model embeddings, hidden states, logits, probabilities, reaction times, costs, or latency measurements.

Similarity

Cosine similarity and cosine distance are computed on L2-normalized hashed text embeddings. Summary files report within-condition and between-condition comparisons. Similarities reflect overlap in the serialized representation text and should be interpreted cautiously.

Projection

2D projections use deterministic PCA via numpy.linalg.svd on centered embedding matrices. Component signs are fixed by forcing the largest absolute component loading to be positive.

Condition-level PCA explained variance ratio: PC1=0.607521, PC2=0.177192
Sliding-window PCA explained variance ratio: PC1=0.495949, PC2=0.147475

Outputs

condition_embeddings.npz
condition_embedding_vectors.csv
condition_embedding_metadata.csv
condition_embedding_similarity_pairs.csv
condition_embedding_projection_2d.csv
sliding_window_embeddings.npz
sliding_window_embedding_metadata.csv
sliding_window_embedding_similarity_summary.csv
sliding_window_embedding_projection_2d.csv
embedding_model_config.json