Marginalia v2

codes as matched filters · token embeddings · 100% in-browser model: not loaded

1 Document demo loaded

Name Source text

2 Codebook 0 codes

Positives define what fires the filter; negatives subtract look-alikes. One example per line. E.g. a person-name code: positives are utterances that contain names ("I spoke with Dr. Ramirez"), negatives are discussion about naming ("we should pick a name for the product").

3 Filter tuning

Window sizes (tokens)

Stride (tokens)

Smoothing (positions)

Negative weight λ

DC removal

Model

First run downloads the embedding model (~25 MB) into browser cache.

→ Signal tracks & detections

enter z ≥ 2.0 exit z ≥ 0.8 min prominence 0.5