Portfolio — Editorial Systems

The Editorial
Engine System

A system that restructures how publishing teams make decisions before translation begins — reducing ambiguity, revision cost, and rework. Built by a translator who refused to be replaced.

AI used here not for generation — for structured analysis, consistency enforcement, and scalable decision-making across complex manuscript inputs.

3 Engines
35+ Output Documents per Run
10+ Years Domain Expertise
Operational — Available for Projects
Source — Hebrew Manuscript
Adapt
Translate
Edit
Target — English Publication
The Background
Editor & Translator
Role: Obsolete
System Designer

Publishing workflows fail before translation begins.

Manuscripts enter translation carrying unresolved structural problems, cultural gaps, and narrative instability. These aren't caught until the translation is already done — when fixing them is expensive, slow, and often incomplete.

After a decade inside publishing houses as a translator and editor, I had seen this failure pattern across dozens of projects. When AI tools arrived and began replacing translators, I didn't retreat. I used that domain knowledge to map every point where the workflow breaks — and built a system to fix them before translation starts.

The Problem

Translation without a system is translation without a safety net.

The workflow that existed when I started building this system had no quality gates, no structure, and no scale — and it treated every manuscript as if it were already ready:

STEP 01
Single locked
ChatGPT prompt
STEP 02
Friction
Manually copy chunks
<1,000 words
STEP 03
Paste into chat.
Get translation.
STEP 04
Paste output into
separate Word doc
STEP 05
Friction
Manually compare
to Hebrew original
STEP 06
Risk
Catch hallucinations
& added content
STEP 07
Repeat for every
chapter

And no one asked: is this manuscript even ready to be translated?

The Insight

The problem isn't translation.
It's the absence of structured decision-making before it.

Every failure I had seen — costly revisions, cultural misfires, inconsistent output — traced back to the same root: decisions that should have been made before translation began were being made during it, or not at all. The solution wasn't a better translation tool. It was a system that restructured when decisions happen.

1
Pre-Translation
Is this manuscript ready to be translated?
Structural problems, cultural gaps, and POV instability compound in translation. The time to find them — and decide what to do about them — is before the work begins.
2
Translation
Can we translate it faithfully, at scale, without losing the original?
Hallucinations, additions, and loss of voice are systematic risks. They require systematic safeguards — not manual checking after the fact.
3
Post-Translation
Is the resulting English actually good English?
A faithful translation can still be full of awkward constructions and rhythm problems. This step exists — and no AI pipeline I have encountered includes it.
Reduces
Revision cycles and rework cost
Standardises
Editorial decisions across manuscripts
Enables
Scaling without adding headcount
Improves
Cross-market success rates
The System

Three Engines. One Pipeline.

Each engine hands off to the next. No stage begins until the previous one is resolved.

Engine 01
Cross-Market Adaptation Engine
A 15-pass diagnostic pipeline that maps every risk in the Hebrew manuscript before translation begins. Includes a mandatory gate that can halt the process entirely.
Pre-Translation
Engine 02
Translation Engine
A 6-layer pipeline that handles the translation with built-in hallucination safeguards, automated sectioning, and structured Word output with editorial comments.
Translation
Engine 03
Line Editing Engine
A 13-pass editorial review of the completed English translation — treating the text as an English manuscript that needs a professional editorial pass.
Post-Translation
System architecture

The full pipeline at a glance

From legibility gate through final editorial pass — every stage, every decision point, every handoff.

The Editorial Engine System — full pipeline diagram Three-engine pipeline: Hebrew manuscript enters a Legibility Gate (pass / conditional / reject), then flows through Engine 01 Cross-Market Adaptation, Engine 02 Translation, and Engine 03 Line Editing, producing publication-ready English. Hebrew manuscript Source text Legibility Gate Pass Reject → author Conditional Proceed with documented caveats Engine 01 Cross-Market Adaptation · 5-pillar diagnostic framework · 15 passes · Composite risk score 0–50 · Adaptation strategy report · Mandatory legibility gate 0–10 21–30 41–50 Adaptation report Engine 02 Translation L1 · Source quality assessment L2 · Whole manuscript analysis L3 · Chapter context packet L4 · Chunk translation + safeguards L5 · Chapter reconstruction L6 · Fidelity audit ← Hallucination check + audit Translated draft Engine 03 Line Editing · 13 discrete passes · Rhythm · Clarity · Voice · POV integrity · Concision · Synthesis + annotated output Publication-ready English Word doc · Editorial comments · Exit packet No translation until manuscript passes gate Single file upload Auto-sectioning No existing pipeline includes this step Full run: under one week vs. month-long manual workflow
Engine 01

Cross-Market Adaptation Engine

"The most expensive mistake in literary translation is starting too early."

Decision Logic — Legibility Gate

Every manuscript runs through the gate before anything else begins. The gate produces one of three outcomes — and each one determines a different downstream path:

✓ Pass
Proceed to full pipeline
All three engines run in sequence. Full diagnostic, translation, and line editing.
⚠ Conditional
Proceed with documented caveats
Risks flagged and recorded. Translation proceeds with specific problem areas marked for human review.
✕ Fail
Return to author. Do not translate.
Manuscript is not translation-ready. Structured feedback report delivered. No resources spent on a text that will require full revision.

Five-Pillar Diagnostic Framework — 15 Passes

Pillar I
Cultural Transposition Strategy
Pillar II
Emotional Legibility
Pillar III
Narrative Coherence & POV Integrity
Pillar IV
Dialogue & Voice Structural Alignment
Pillar V
Pacing Propulsion

Composite Risk Score — 0 to 50

Export Ready
0–10
Minor Adaptation
11–20
Moderate Risk
21–30
High Adaptation
31–40
Major Intervention
41–50

Sample Chapter Diagnostic Output

What a chapter-level diagnostic report looks like in practice — the structured output that editors and publishers receive before any translation work begins:

Chapter 3 — Diagnostic Report
Emotional clarity
Low
Protagonist's internal state shifts without transition. Reader loses emotional anchor at chapter midpoint.
Dialogue alignment
Moderate
Voice registers inconsistent between characters. Distinguishable in source; risk of flattening in translation.
Cultural transfer risk
High
Three idiomatic references with no English equivalent. Adaptation strategy required before translation begins.
POV integrity
Stable
Consistent third-person limited throughout. No unexpected perspective shifts detected.
Gate recommendation
Conditional — proceed with adaptation strategy for cultural references documented in sections 2.3 and 2.7. Flag emotional clarity issue for author review before translation of chapter midpoint.
Engine 02

Translation Engine

Six structured layers — from pre-translation quality assessment through final fidelity audit. The original Hebrew file is uploaded once; the engine handles the rest.

L1
Source Quality Assessment — 10-Axis Diagnostic
L2
Whole Manuscript Analysis
L3
Chapter Context Packet
L4
Chunk Translation — with Safeguards
L5
Chapter Reconstruction
L6
Fidelity Audit
Key Innovations vs. Previous Workflow
Single File Upload
The original manuscript is posted once. No manual copy-pasting of sub-1,000-word chunks.
Automated Sectioning
The engine works through the text in optimized sections automatically. The translator does not manage segmentation.
Hallucination Safeguards
Built-in protocols detect and flag content not present in the source — rather than relying on manual comparison.
Structured Word Output
Complete translation delivered as a Word document with inline editorial comments where review is needed.
Engine 03

Line Editing Engine

The difference between a translation that is faithful and a translation that is good English. Thirteen discrete passes — each addressing a specific editorial dimension, in sequence.

PASS 01
Mechanical Integrity
PASS 02
Clarity & Agency
PASS 03
Modifier Attachment
PASS 04
Concision
PASS 05
Sentence Energy
PASS 06
Rhythm & Readability
PASS 07
Dialogue Clarity
PASS 08
POV Integrity
PASS 09
Continuity
PASS 10
Show vs. Tell
PASS 11
Exposition Density
PASS 12
Phrasing & Word Choice
PASS 13
Repetition & Echo — followed by Synthesis & Annotated Output
"A post-translation editorial pass applied as a structured AI pipeline — a step I have not encountered in any existing translation workflow."
In Practice

Tested Across Multiple Manuscripts — Ready for Use

The pipeline has been developed and refined through multiple runs on Hebrew literary fiction manuscripts. Each run produced structured output across all three engines — from Legibility Triage through final Line Editing. The system is operational and ready for deployment on new projects.

3
Engines, Full Pipeline
35+
Output Documents per Run
v8
Current Engine Version
After Multiple Calibration Runs
SQA Output
AVC Checkpoints
Chapter Translations
Fidelity Audit
LEE Annotated Output
Exit Packet
The Transformation

What Changed

The system changes the sequence — and that changes the timeline. By directing the editor's attention to exact problem locations, completing the structural grunt work automatically, and eliminating the need for separate uploads and manual cross-referencing across tools, what previously took close to a month can be completed in less than a week.

Before

No quality gate — manuscript entered translation regardless of readiness
Manual copy-paste segmentation, capped at 1,000 words per chunk
Hallucinations caught manually, after the fact, by comparing output to source
No post-translation editorial pass — rough English left in the output
Structural problems in source text discovered mid-translation
Timeline: weeks to a full month per manuscript

After

Mandatory Legibility Gate — manuscripts that aren't ready don't proceed
Single file upload; automated sectioning managed by the engine
Built-in hallucination safeguards and fidelity audit in every run
13-pass line editing engine delivers publication-ready English
Cross-market diagnostic maps all structural risk before translation begins
Timeline: the same manuscript in less than a week
Built Through Iteration

From a Single Prompt to a Three-Engine System

Each version was tested against real manuscripts. What broke shaped what came next. Here is what actually broke.

Origin
One ChatGPT Prompt
A single locked prompt. Manual copy-paste under 1,000 words. No safeguards, no gates, no structure. The starting point that made every failure visible.
Adaptation Engine — v7 → v8
Structural Refinement
v8 added Pass 1b, Pass 10b, Pass 10.5, a Re-Run Protocol, and a standalone Legibility Triage gate — separated from the engine to allow independent updates without version conflict.
Current — Operational
Three-Engine Pipeline
Cross-Market Adaptation Engine v8 · Translation Engine v1.1 · Line Editing Engine. Tested across multiple Hebrew literary manuscripts. Ready for deployment.
What actually broke — and what it forced
Failure
L4 chunks were hallucinating additions
The translation layer was generating content not present in the Hebrew source — dialogue tags, filler phrases, soft expansions. Manual comparison was catching some of it. Not all of it. The problem was systematic, not occasional.
What it forced
L6 Fidelity Audit, built from scratch
A dedicated sixth layer that cross-references translation output against source structure — at the level of sentence count, paragraph breaks, and flagged additions. Manual post-hoc checking eliminated entirely.
Failure
Diagnostics were blocking valid manuscripts
Early versions of the Adaptation Engine ran a single diagnostic pass that produced false positives — flagging manuscripts as high-risk based on surface features rather than structural ones. Good books were being returned to authors unnecessarily.
What it forced
Legibility Triage separated from the engine
The gate became a standalone pre-pass — with its own scoring criteria, conditional paths, and documented caveats. It runs before the engine loads and can be updated independently without version-conflicting the full pipeline.
"
This is not AI replacing a translator.
It is a translator who refused to be replaced —
and built something better.

The Editorial Engine System is operational and available for deployment on new projects —
for publishing houses, literary agencies, and translators working with Hebrew-language fiction.

Get in Touch

Interested in the pipeline?
Let's talk.

adi.d.kafri@gmail.com
🔗 linkedin.com/in/adi-kafri
📄 Download CV / Résumé
Translation & writing excerpts available on request
Operational — Available Now