Why we refuse to fill in missing text
Every generative model we tested would happily write a paragraph into a hole in a will. Here is what document restoration should do instead - and why refusing to invent text is the harder product decision.
The Docovly team
Document restoration
Read back & checked
The first serious test we ran was a 1954 will with a torn-away bottom corner. Roughly forty words were physically gone - not faded, not obscured, gone with the paper. We handed it to five restoration models and asked each to restore the page.
All five returned a complete document. Four of them produced text in the gap that was grammatically perfect, stylistically consistent with the surrounding clauses, and entirely invented. The fifth produced a smudge. The smudge was the only honest answer in the set.
Plausible is not the same as correct
A generative model asked to complete a document is doing exactly what it was built to do: producing the most likely continuation. On a landscape photograph, the most likely continuation is fine - nobody is harmed by an invented leaf. On a testamentary document, the most likely continuation is a forgery with excellent grammar.
The uncomfortable part is that the invented text was good. It read like the rest of the will. A solicitor skimming the restored page would have had no signal that those forty words came from a model rather than from the deceased. That is the whole problem stated in one sentence: the danger of a generative restoration is not that it fails visibly, it is that it succeeds invisibly.
The failure mode is not that the model produces nonsense. It is that the model produces something indistinguishable from the truth.
Where the line actually falls
The distinction we settled on is whether the missing content is determined by what survives. Some of it is genuinely recoverable, because the surrounding material constrains it completely. Some of it is not, because the only evidence for it was on a piece of paper that no longer exists. The entire discipline of honest restoration lives on that line, so it is worth being exact about which side each kind of damage sits on.
What survives determines - so we rebuild it
- Paper texture and tone - fully determined by the sheet around the loss, so a missing patch of substrate can be rebuilt without inventing anything.
- Rules, borders and table gridlines - their geometry continues from both sides of the gap, so the line that crossed the hole is the line that met it on either edge.
- Guilloche and security patterns - periodic by design, so the pattern is recoverable from its own repetition elsewhere on the page.
- Letterforms partially crossed by a tear - the stroke evidence survives on both sides of the break, and the letter is reassembled from the strokes that remain, not guessed from context.
What survives does not determine - so we mark it
Text that existed only on paper which is now gone is not recoverable by any means, and no amount of surrounding context makes it so. Context can make a guess more convincing; it cannot make it true. That text is recorded as a loss, with its extent measured, and rendered as an overlay if the archive wants to show exactly what is missing and how much.
The test we apply to every pixel
Would this survive a second, independent operator arriving at the same result from the same evidence? Rebuilding a gridline passes. Writing a plausible clause into a hole does not - two honest operators would write two different clauses, which is proof the evidence never determined it.
Reconstruction, without invention
It would be easy to read all of this as "the tool does not reconstruct anything," and that is wrong. Reconstruction is used heavily. Torn certificates come back with their borders whole. Punched and taped ledgers come back with their rules continuous. What reconstruction will not do is cross the line from geometry into language - it rebuilds the things the page proves, and stops at the things the page merely suggests.
| Damage | Recoverable from what survives? | What we do |
|---|---|---|
| Torn corner, blank paper | Yes - substrate is uniform | Rebuild the paper |
| Gridline crossing a hole | Yes - geometry continues | Reconnect the line |
| Half a letter lost to a tear | Yes - stroke evidence on both sides | Reassemble the letterform |
| A whole word lost with the paper | No - only context remains | Mark as loss, never fill |
| Faded ink, still present | Yes - the ink is there, just weak | Restore stroke density |
This costs us demos, and we accept that
We are aware of how this looks in a side-by-side comparison. A competitor's output is complete and beautiful. Ours has a marked rectangle in the corner with a note saying roughly forty words are unrecoverable. In a thirty-second evaluation, theirs wins every time, because the evaluator is comparing two pictures and one of them has a hole in it.
In an archive, a registry or a court, ours is the only one of the two that can be used. A restored will that might contain invented clauses is not a restored will - it is a new document of unknown authorship. The marked loss is not a weakness in the output; it is the single most important thing the output tells you.
A restoration you cannot trust is not a restoration. It is a very confident guess wearing the clothes of the original.
What this means if you are choosing a tool
If you are evaluating document restoration software for anything with legal, financial or historical weight, the demo to ask for is not the beautiful before-and-after. It is the damaged page with a genuine loss in it. Watch what the tool does with the part it cannot possibly know. A tool that fills the hole is telling you, clearly, that it will do the same thing to your records the first time you are not looking.
Reconstruction is available and used heavily - on paper, rules, borders, security patterns and partial letterforms. What it will not do is generate words to fill a gap. See the Document Reconstruction page for the full boundary, worked through example by example.
Keep reading
“A blank should mean the document was blank.”
Anything else deserves its own column
The withheld column, and why a blank is not a blank
Structured condition, not a paragraph
What a conservator taught us about condition surveys
2-3 years - From printed to blank, faster than ink ever fades
Thermal receipts are a time bomb, and most of them have already gone off