Work/03 · Editor Pipeline

The Editor Pipeline

A 7-stage editor that strips AI-generated patterns and rebuilds a writer's authentic voice — in English and 中文.

Voice-profile modeling Test-driven eval harness EN / ZH bilingual Deterministic checks
24EN pattern detectors
21ZH pattern detectors
7pipeline stages
1voice profile per writer
Proof exhibit · the signature

Drag to rebuild the voice.

Left is an AI-slop draft. Drag the handle right and the pipeline's edit takes over — struck-through patterns are what 24 detectors flagged and removed.

Removed AI pattern Rebuilt in Li's voice ← drag the handle →
AI draft Li's voice

In today's fast-paced world, it's important to note that writing with AI can unlock a plethora of opportunities. Whether you're a seasoned professional or just starting out, these tools empower you to elevate your content and take it to the next level. In conclusion, the possibilities are truly endless.

AI changes what one writer can do in an afternoon. Used well, it's a faster draft and a sharper second opinion — not a replacement for having something to say. The tool is loud; the voice has to be louder. That's the whole job: keep the person, lose the autopilot.

Hedging openers

Cuts it's important to note, in today's world — filler that delays the point.

Inflated diction

Flags plethora, elevate, unlock — thesaurus words a person wouldn't say aloud.

Empty symmetry

Removes whether you're…or… and take it to the next level boilerplate.


Proof exhibit · a published essay

A real edit, draft to published.

This is one of my own essays — “Procrastination is a three-dimensional problem.” Left is the raw draft, right is what published: same argument, the slack cut out.

原稿 · raw draft42 字

拖延,与完美主义没有关联。拖延是大脑在专注于要做事情要付出的成本与努力,而不是可期待的回报。

发表稿 · published21 字

拖延跟完美主义的关系,远没有人们以为的那么大。

拖延真正变得棘手,是两个维度同时起作用:时间不确定性
The thesis in one line — exactly the kind of sentence the pipeline exists to protect.

Pipeline

Seven stages, voice intact.

01

Voice profile

Model the writer's cadence, vocabulary, and tics from a held-out corpus of their real work — the reference everything else is measured against.

02

Pattern detection

Run 24 EN + 21 ZH detectors for AI tells: hedged openers, inflated diction, empty symmetry, em-dash overuse, and listy "tricolon" padding.

03

Strip

Remove flagged spans without collapsing meaning — the diff you just dragged is this stage made visible.

04

Rebuild

Re-draft the stripped passage toward the voice profile — restore rhythm, restore the argument, restore the person.

05

Bilingual pass

For 中文, apply ZH-specific detectors — translationese, over-formal connectors, and AI's characteristic flat parallelism.

06

Eval harness

Score the result against deterministic checks and a cross-family LLM judge on a held-out set — no edit ships that regresses voice fidelity.

07

Diff & report

Emit a reviewable diff with every removed pattern labelled — the writer stays in control of the final call.

Rigor

Edited like software is tested.

Held-out corpus

Voice fidelity is scored against samples the model never saw during profiling — so "it sounds like Li" is measured, not asserted.

Deterministic checks

Pattern counts, sentence-length variance, and forbidden-phrase lists run as pass/fail gates — repeatable on every run.

Cross-family judge

A model from a different family grades the edit, so the system can't just flatter its own style.

Regression guard

No edit ships if it lowers voice fidelity below the prior release — the harness blocks it, the same way a failing test blocks a deploy.