Orders open by email · Reply within 1 business day. Questions? maciek@studiochrome.ai

The standard behind every source

The record first.
The conclusion second.

A convincing transcription can still contain the wrong name. We use AI to help read historical handwriting, then check the facts and the links between people.

What we measured

Useful at reading.
Not safe unchecked.

Our internal benchmark found an estimated 9% raw AI misreading rate in name fields. That is why unchecked AI names do not belong in a delivered evidence file.

48

acts in the full benchmark

31 Polish, 16 Russian and 1 Latin record. German handwriting was not tested.

≈9%

estimated raw name misreadings

After separating some spelling and indexing conventions from the 21 normalized mismatches. This measures the AI draft, not a verified report.

0 / 2

pilot errors caught by a second AI

On the earlier 13-act pilot, both models made the same two residual errors. Agreement between models was not independent verification.

Internal T1 benchmark, full-run verdict · 22 October 2026

What the number means. This is a small internal study, not an independent audit or a final-report accuracy score. Human index rows supplied the comparison labels and can themselves contain mistakes. One Latin act and no German acts cannot establish accuracy for those scripts. We have not established a measured error rate for completed, human-checked reports.

Record languageActs testedNormalized name mismatchesHow to read this
Polish3117 / 124 fields · 13.7%Includes some differences in spelling and indexing conventions.
Russian164 / 42 fields · 9.5%A small test of handwritten Russian records, not every script or archive.
Latin10 / 4 fieldsToo little evidence to claim a Latin accuracy rate.
German handwriting0Not testedNo benchmark claim for Kurrent or Sütterlin.

The approximately 9% estimate separates likely misreadings from some of the 21 normalized mismatches above. Dates, places and final-report accuracy were not measured by these name-field results.

Five steps from scan to evidence.

Locate the original

Record the archive, parish, book, year, act number and scan link. Keep a crop of the actual entry, so the reading can be checked again.

Read each field

Draft the original-language transcription and English translation. Preserve historic spellings. Mark an uncertain letter or missing word instead of silently filling it in.

Check independently

Compare indexed names field by field with the human-created index. A person returns to the scan when the two disagree. An index is a useful check, not an infallible source.

Test the family link

Compare parents, places, ages, occupations and other identifying details across records. Explain why two entries refer to the same person, and keep an inferred link separate from a directly supported one.

Show the limits

Deliver the sources, unresolved readings, probable links and searches that found nothing. State which missing document could settle the next question.

Two AI readings are not independent proof. In the earlier 13-act pilot, a second model repeated both residual errors made by the first. Agreement alone did not make those names right.

One unclear word should stay one unclear word.

Reading confidence belongs to individual fields. A legible date does not make a faint surname reliable, and a confident model is not a substitute for checking the scan.

We keep the original reading, any uncertain characters, the check performed and unresolved disagreements together. These are evidence notes, not statistically calibrated percentages.

Sample fieldReading / confidence noteEvidence
Given nameKarol · clear reading; checked by the owner.1828 birth act
Mother’s surnameDonaygrodzka · historic spelling preserved; later records use variants.1852 marriage act
Priest’s signaturePartly unreadable · brackets remain in the transcription; no invented completion.1828 transcription
Krystyna’s given names“Izydora Krystyna” in the reading, “Krystyna” in the index · still flagged for review.1832 birth act

Missing evidence is part of the result.

Where no human index exists, a person must check the reading. Low confidence requires closer attention; high confidence does not remove that requirement. Unindexed work is scoped as an AI draft checked by a person, with the additional review time made clear.

A faint scan, a missing volume or a contradictory birthplace may remain unresolved. We record what was searched, what was ruled out and what would change the conclusion. We do not replace a missing act with a plausible story.

The sample includes a death record that says Magnuszew while the earlier birth and marriage support Rozniszew. The conflict is retained in the record notes, so the reader can examine the evidence.

Evidence for the legal question.

Our work finds and explains the records. A birth act can prove a family fact; citizenship also depends on acquisition, later events and the law that applied at the time. A qualified legal reviewer assesses complex citizenship chains, filing and appeals.

Scan crops and research translations are working evidence. An authority may still require certified copies, apostilles or sworn translations. A well-supported family tree is not a promise of citizenship.