A research prototype for meaning-preserving text transformation with semantic and quality evaluation.
中文说明 · Product brief · Architecture · Research notes
ScholAI explores how a text-transformation pipeline can preserve claims, named entities, and numerical facts while evaluating semantic consistency and language quality. It is an experimental project, not a production service.
The repository combines controlled local-model generation with layered evaluation:
- Extract claims and constraints from the source text.
- Generate candidate rewrites with controlled sampling.
- Check semantic consistency with embeddings and NLI.
- Evaluate quality using statistical and model-based signals.
- Retry or reject candidates that miss the configured quality bar.
This project is intended for legitimate editing, accessibility, localization, and writing-assistance research. It must not be used to misrepresent authorship, evade academic-integrity requirements, or bypass platform safeguards. Any deployment needs its own user policy, human review, and compliance assessment.
Research prototype. The repository contains product documentation, architecture notes, experiments, and tests; interfaces and evaluation thresholds may change as the research evolves.
# Create the environment and install dependencies.
uv sync
# Inspect available development tasks.
make helpSee the project documents above for model requirements, evaluation design, and hardware guidance. Keep credentials in local environment files; never commit API keys.
ScholAI 是一个“语义不走样、质量可检查”的文本转换研究原型。它通过受控生成、语义一致性验证和质量评测来探索文本编辑工作流；不是生产服务，也不应用于代写、学术不端或规避平台规则。
The repository includes focused agent roles under .claude/agents/ for engineering, product review, and QA critique. Treat generated outputs as drafts: validate meaning, facts, and policy compliance before use.