A Transparent Checker for Stiff AI Writing

A Hugging Face repository identified as erikkannike/qwen3-4b-deslop-v4 recorded a commit on Oct. 18, 2026. The listing places that identifier and revision history on Hugging Face. Hugging Face

A second repository identified as erikkannike/qwen3-4b-deslop is also listed on Hugging Face. Hugging Face Its card describes its slop score as a regex-based detector of AI writing tropes reported per 1000 words. A regex detector searches for fixed phrases using set rules, not AI judgment. The description states that ordinary human writing scores close to zero on the slop score.

Qwen has announced the release of Qwen3 as the latest addition to the Qwen family of large language models. Qwen That announcement mentions Qwen3-4B.

The design choices shape how it can be used. The detector is deterministic and inspectable. Its patterns can be read, tested, and patched without retraining a model. Precision can be high for known phrases. Recall is limited to what is on the list. Reworded text evades it, and users trying to bypass it can do so quickly.

Reporting per 1000 words is standard practice for style telemetry, or routine measurement of writing style. It allows comparison across prompts, completions, and checkpoints. Raw counts would penalize longer text. Rates show tendency. If the near-zero baseline for ordinary human writing holds in a given deployment, operators have a calibration anchor. It defines what quiet looks like. That simplifies setting thresholds, sending alerts, and A/B comparison of decoding settings or post-processing.

In my view, the engineering question is where such a scorer sits in the stack. Used offline, it can audit sampled outputs, compare system prompts, or gate dataset construction. Used inline, it can resample, rewrite, or flag. Latency is negligible. Cost is negligible. The failure mode is not performance. It is drift. Tropes change. Model habits change. A static pattern list decays unless maintained. Worth flagging for teams that would operationalize this kind of signal: version the patterns with the same discipline as model weights, log the scorer version alongside the generator version, and track false-positive rates on human-authored controls.

The broader context here is unglamorous and useful. Style control has become infrastructure work. Teams tuning assistants for enterprise or developer use need repeatable ways to measure woodenness without calling a large judge model for every batch. A small, transparent heuristic does not replace human review or model-based evaluation. It reduces how often those expensive checks must run. Over time, that kind of tooling tends to compound. Small filters make larger systems shippable.