Table 2 · pipeline
Load, paraphrase, split, stack, classify.
The Colab method, running here on CPU-class sparse models. GPU is only needed for a live rewrite in the probe. A document and its rewrite always share a split.
- 01queued
Load
proxy generator
raw table
- 02queued
Clean
regex, token floor
n_tok ≥ 40
- 03queued
Synthesize
polish-A / rhythm-B
coda_synth
- 04queued
Id-safe split
stratified on doc_id
70 / 15 / 15
- 05queued
Features
TfidfVectorizer, stylo
sparse + scaled
- 06queued
Classify
LinearSVC, logistic
two models
- 07queued
Ablate
word / +char / +stylo
three F1s
- 08queued
Evaluate
held-out test
reports + plots
Leakage control
A document and its rewrite share content. If one is in train and the other in test, the classifier cheats on topic words. We split on original doc_id first (70 / 15 / 15), stratify by topic, then attach each id’s human and llm rows to the same split.
Rewrite prompt
Rewrite the following document in a different wording and sentence rhythm.
Keep the same facts, names that are already masked, and the same topic.
Do not add products, prices, instructions, or new claims.
Do not mention that you are an AI.
DOCUMENT:
{text}