II.7 Quality model: lossless, lossy, recognised, generated
The question behind every app
Before you press a button, ask: what does this app do to my content? There are four honest answers. The original book plan had three; reading the app chapters showed a fourth, in between, that deserves its own name.
| Class | What it does to your content | Can it surprise you? | Can you undo it? |
|---|---|---|---|
| Lossless | Keeps content as it was; changes arrangement, access or wrapping | No | Usually, from the original |
| Lossy | Throws detail away or re-draws pages | Visibly, yes | Only by going back to the original |
| Recognised | Estimates text from pixels (OCR) | Quietly: wrong words that look right | Re-run with better input |
| Generated | A model writes or decides something new | Yes: it varies from run to run | Run again; always review |
Exact <-----------------------------------------------> Invented
lossless lossy recognised generated
(pages (softer, (estimated (written by
copied) re-drawn) text) a model)
Class 1: Lossless
The content is carried over unchanged. The app only rearranges, adds, removes or wraps.
Splitter (pages are copied, not re-rendered), Delete Page, Rotation at 90, 180 and 270 degrees, Booklet (pages placed, not re-drawn), Merger for PDF plus PDF, Add TOC, Add Table (adds pages), Flatten (same look, no longer interactive), Protector in full document lock, Unlock, Sign, Optimize (by the site's own statement), Resize in standard mode (text stays text).
Read-only apps that never change a page also sit here: the Extract apps, Parser, PDF/A Validator, Verify Signature, and Search (adds highlights; on a scan its recognition step applies).
"Lossless" does not mean "harmless". Flatten is one-way; signing makes later edits detectable; Add Table inserts pages that shift page numbers.
Class 2: Lossy
Information is discarded or the page is rebuilt.
| App | What is lost |
|---|---|
| Compressor | Picture detail, on purpose (JPEG, lower resolution). Text and vector art stay sharp |
| Deskew | Pages are rebuilt as images; do not count on a text layer surviving |
| Resize tile mode | Pages become images; the file may grow; text is only searchable with OCR |
| Rotation Auto Skew | Content is resampled |
| Unwatermark | A PDF comes back as plain re-typeset text |
| Recovery | Only where damage forced a rebuild; the report says which pages |
| Converter | Depends on target: Word loses exact layout, Excel is not a page picture, images lose text |
| Collage | Photos kept as JPEG at slightly reduced quality |
Rule: never push a lossy result through the same lossy step again. Compress once, from the original.
Class 3: Recognised (an estimate)
OCR reads pixels and produces text. It is software, not AI writing, but it is an estimate, and a wrong word can look entirely plausible.
OCR Image, Searchable, Merger with its OCR switch, Converter with OCR, Search and Abstracts on scanned input. Searchable also offers an AI correction pass that proofreads the recognised text; that pass is generated and can change words.
Habits: set the language (II.6), straighten first, and spot-check names, numbers and legal wording. "No results" in a search of a scan does not prove absence.
Class 4: Generated (a draft)
A model decides or writes. Output varies from run to run, and it can contain mistakes or invented details.
| Family | Apps |
|---|---|
| Add & Structure | Table of Contents AI (outline, with a rougher rule-based fallback), AI Illustrator, Collage layout |
| Convert & Create | SVG Maker, Chat |
| Extract | Resume Analyzer |
| AI understand | Ask PDF, PDF Knowledge Base, Abstracts, Cheat Sheet, Checklist |
| Developer apps | Assembly Chat, Repo Docs, Refactor Assistant, Page SEO Checker, Advisors and Document Engineer |
Some of these include built-in safeguards: Ask PDF cites pages, and Abstracts checks non-fiction claims against the document. They reduce risk; none promises correctness (Trust). Hand-built twins exist for several: Add TOC for TOC AI.
Why this matters for order in a chain
The class of each step predicts what it will damage for the next step.
- Lossless work first. Fix pages, order and orientation while content is untouched (Organize).
- Recognise before you analyse. OCR before Search, Extract Text or AI apps, because they need a text layer. Deskew and rotate before OCR.
- Generate on clean text. AI apps are only as good as their input; garbage recognition becomes confident garbage.
- Lossy steps late, and once. Compress after editing.
- Signatures and passwords last. Almost every step above changes the file bytes and breaks a signature. Sign last; protect near the end.
- Generated output is a draft. Do not feed it back as source material for the next AI step without reading it.
Part III turns this into the golden rules of ordering and a dependency matrix.
A mini checklist before pressing the button
- What class is this app? (Look at "Quality & limits" in its chapter.)
- Is my original safe? (Originals are never modified, but keep a copy anyway.)
- Is this step lossy? Then it should be the last of its kind.
- Is the output a draft? Then plan a review.
Try it
Pick a scanned PDF. Run Searchable, then Compressor, then compare to doing it the other way round. You will see why the order is a quality decision, not a style choice.
Next
Part II is done. Continue to the map of how the apps relate.