Show HN: Slopsift – a local, graph-backed linter for AI writing

SlopSift, a local graph-backed linter for AI writing, uses a custom-trained dependency parser to detect canned arguments, unsupported claims, and filler in text. The tool runs on-device via quantized ONNX weights in Node and browser WebAssembly, avoiding API calls. It outputs findings in ESLint-shaped JSON for CI integration.

Small enough to ship Quantized ONNX weights run locally in Node and browser WebAssembly. Natural-language processing. Local first. SlopSift uses a small dependency parser we trained to map the relationships between words and find canned arguments, unsupported claims, and filler. Loading the on-device NLP model… More than a word list Many writing tools stop at word matching and basic parts of speech. SlopSift follows the relationships between words, so rules can recognize how a sentence makes its claim—not only which vocabulary it uses. Our custom-trained compact model maps tokens, parts of speech, and the grammatical relationships holding the sentence together. Authorable rules inspect the graph for structural tells. Every finding names what matched and the exact text that triggered it. Errors are strong tells. Warnings need attention. Notes are candidates—not a machine pretending to know who wrote the sentence. Not an API wrapper SlopSift starts with a compact pretrained English encoder and trains it for parts of speech and dependency parsing. Training combines structured distillation from a larger parser with 50 controlled examples targeting grammatical relationships used by the linter. We reserved separate template families for evaluation. Quantized ONNX weights run locally in Node and browser WebAssembly. UPOS, dependency arcs, and dependency relations—not a generic text score. The model and deterministic rules run on-device. No remote judge reads the text. “As an AI language model...” Three paragraphs use the same canned outline. An actorless passive may be hiding responsibility. Glob files, lint Markdown, inspect code comments, and emit ESLint-shaped JSON in CI. Let coding agents run the real linter, interpret its findings, and edit without flattening your voice. Yes, AI helped build this. That is the point. SlopSift is not an AI detector and it does not pretend to know who typed a sentence. It catches vague or inflated writing. It also catches repetition and borrowed certainty. Human beings do those things too. Use --format json for machines, --level info for the full suspicious pile.