Skip to content

pipeline.md

Latest commit

 

History

History
84 lines (59 loc) · 4.46 KB

pipeline.md

File metadata and controls

84 lines (59 loc) · 4.46 KB
Raw

Standard Pipeline (v1.5.1) — Production Path

Positioning: The Standard Pipeline is the production-grade integration added in v1.5 and tuned in v1.5.1 against real samples. It combines Method 1 (Translation Chain) and Method 2 (LLM Rewriting) from the 4 methodologies into a fixed, validated 4-step chain. This is the recommended path for actual use.

See examples/showcase/ for 5 end-to-end traces of every intermediate step on real input texts.

Architecture

Input Text (EN)
    ↓
Step 1: LLM (temp 1.3)                    ── 中文改写
    Input → Chinese + Humanization Rewrite
    ↓
Step 2: LLM (temp 1.3, with history)       ── 日语改写
    Chinese → Japanese + Humanization Rewrite
    ↓
Step 3: Google Translate                    ── 一轮翻译
    Japanese → Finnish
    ↓
Step 4: Niutrans                            ── 二轮翻译
    Finnish → Target Language (EN)
    ↓
Output (Humanized EN)

Steps 1–2 call any OpenAI-compatible chat API. The default provider is DeepSeek; set [llm].provider = "openrouter" to route through OpenRouter. See configuration.md.

Why Each Step Matters

Steps 1-2: LLM Humanization Rewrite

These steps do the heavy lifting. The configured LLM at temperature 1.3 doesn't just translate — it rewrites. The key differences from plain translation:

  • Sentence restructuring: AI-typical uniform sentence patterns get broken
  • Vocabulary diversification: Formal/robotic word choices get replaced with natural alternatives
  • Rhythm variation: The output has varied sentence lengths (burstiness)

Step 2 carries the conversation history from Step 1. This gives the LLM context about what was already changed, preventing it from reverting patterns that Step 1 disrupted.

Steps 3-4: Cross-Engine Translation Chain

Two translation hops through two different engines compound structural changes:

  • Google (Step 3): Neural machine translation with the largest training corpus, applied to the Japanese → Finnish hop
  • Niutrans (Step 4): Different NMT architecture and training data, applied to the Finnish → English hop

Using different engines prevents any single-engine fingerprint from surviving. Each engine restructures grammar differently, and the cumulative effect produces text that doesn't match any known AI generation pattern.

Language Distance Strategy

The chain maximizes linguistic distance at each hop:

Hop Languages Distance
1 English → Chinese High (different family, no shared script)
2 Chinese → Japanese Medium (shared characters, different grammar)
3 Japanese → Finnish Very High (Japonic → Uralic, SOV → SVO, agglutinative)
4 Finnish → English High (Uralic → Germanic)

Finnish was selected for the intermediate step because of its agglutinative morphology — it forces deep restructuring of word forms and clause boundaries, which is hard to reverse-engineer back into AI-typical patterns.

Parameters

Parameter Value Why
LLM provider deepseek (default), openrouter, or atlascloud Set via [llm].provider in config.toml. All use OpenAI-compatible /chat/completions.
Temperature 1.3 Higher than default (1.0) to increase creative variation. Too high (>1.5) causes incoherence.
Model Provider default or [llm].model deepseek-chat (DeepSeek), deepseek/deepseek-chat (OpenRouter), or qwen/qwen3.5-flash (Atlas Cloud). Any compatible model slug works.
Base URL Provider default or [llm].base_url Override to point at a custom OpenAI-compatible proxy.
History 1 round Step 2 sees Step 1's context. More rounds didn't improve quality in testing.
Intermediate language fi (Finnish) Configurable via [pipeline].intermediate_lang in config.toml.

Validation

We ran the pipeline end-to-end on 5 input texts across diverse topics (quantum computing, supply chains, financial literacy, peer review, etc.) and saved every intermediate step output.

Results: All 5 final outputs were classified as human by the AI detector. Confidence scores ranged from 0.7218 to 0.9997.

See examples/showcase/ for the full traces.

Want more tiers? Lynote.ai adds Advanced (multi-round LLM) and Focus (detection-guided feedback loop) tiers on top of Standard.