cx-pangram

Local AI-edit detection with Pangram Labs' EditLens models (Open Pangram, ICLR 2026).

Install with uv tool install "cx-pangram @ git+https://github.com/jmpaz/cx-pangram".

Usage

The EditLens models are gated on Hugging Face. Request access through the Open Pangram collection, then run hf auth login or set HF_TOKEN.

Scores run from 0 (human) to 1 (AI-generated), with chunk-level attribution for longer texts.

Longer texts are scored in roughly 350-word chunks; inputs below 50 words are marked unreliable

CLI

cx-pangram essay.md                            # any local file (or stdin / --text)
cx-pangram score essay.md --chunks             # per-chunk attribution
cx-pangram diff samples/*.txt                  # side-by-side with deltas
cx-pangram score -q --fail-over 0.5 posts/*.md # gate: exit 3 when over
cx-pangram --json ...                          # json / jsonl / md via -f
cx-pangram eval -n 200 --markdown              # faithfulness vs pangram/editlens_iclr

Exit codes: 0 ok · 1 error · 2 usage · 3 gate tripped · 4 gate indeterminate. Unreliable inputs are excluded from gates.

Python

from cx_pangram import get_engine

det = get_engine().detect(open("samples/human_ishiguro.txt").read())
print(det.score, det.band, det.confidence, det.chunks)

get_engine() reuses loaded models. Use detect_batch to score multiple documents together.

Models

key	checkpoint	selected by default
llama	pangram/editlens_Llama-3.2-3B	CUDA or MPS
roberta	pangram/editlens_roberta-large	CPU

Devices are selected in cuda > mps > cpu order; use --device to override. Install the [cuda] extra for 4-bit Llama on NVIDIA, [contextualize] to score external references, or [eval] for the evaluation harness.