When OCR changes the script,
does it preserve the number?

A number has a value and a written form. This controlled evaluation measures whether OCR preserves each one — and makes its evidence, corrections and limitations inspectable.

Research led by Syamjith NK, with disclosed AI assistance.

Read the preprint Explore code & data DOI & permanent archive Download the research package

Same value is not the same script.

These three illustrative records come from archived model outputs. The input strings below are reconstructed displays: the original test images were not supplied. No new inference was run. These examples are not a representative sample or a model ranking.

ARCHIVED EXAMPLE · NEUTRAL PROMPT

Reconstructed input٣٣

Archived output٣٣

Value: correct
Script: preserved

Source record

Model: Arabic-English-handwritten-OCR-v3. Core, Arabic-Indic, neutral prompt.

ARCHIVED EXAMPLE · NEUTRAL PROMPT

Reconstructed input٤٤

Archived output44

Value: correct
Script: changed

Source record

Model: Arabic-English-handwritten-OCR-v3. Core, Arabic-Indic, neutral prompt.

ARCHIVED EXAMPLE · NEUTRAL PROMPT

Reconstructed input٩٥

Archived output90

Value: incorrect
Script: changed

Source record

Model: Arabic-English-handwritten-OCR-v3. Core, Arabic-Indic, neutral prompt.

Download the exact example records

A 30-second explanation

Silent video with on-screen captions. Its examples and conclusions are also available as text on this page.

What the study observed

5 models4,000 core + 3,000 robustness trials

41 / 1,500Neutral prompt: 2.73% joint success on non-Western core trials

33 / 1,500Preservation prompt: 2.20% joint success on non-Western core trials

Joint success requires both numeric correctness and preservation of the input numeral script. Conditional script fidelity uses only outputs that emit decimal digits. A Unicode scoring correction changed 22 categories across the 7,000 archived trials.

What it does not establish

Clean synthetic numeral probes are not production document error rates.
The adjusted H1 observed-cluster difference is −17.25 percentage points, but bootstrap calibration was inadequate. A calibrated H1 significance claim and an effect beyond 10 points are withheld.
Formal H2 equivalence is also withheld. A small estimated difference does not establish equivalence.
H1 selects on output script and is observational. It does not establish that script conversion causes numeric errors.
Original model revision hashes, font hashes and rendered stimuli are missing; exact historical image-level reproduction is qualified.

Open enough to check.

The release includes unchanged raw records, the frozen protocol, corrected derived scores, analysis code, stored bootstrap draws, calibration diagnostics, a change history and a SHA-256 manifest. Analysis can be rerun without calling an OCR service.

Syamjith NK led and directed the research. Author-operated agents, Claude and ChatGPT (called Astra by the author) assisted with code, analysis and writing; Codex assisted the revision and release. Exact original tool versions and task allocation are unavailable.

Code: MIT. Non-software research materials: CC BY 4.0. Third-party rights remain with their holders. A public release is not peer review or conference acceptance.

Discuss or report a reproduction issue Research questions

When OCR changes the script,does it preserve the number?

Same value is not the same script.

A 30-second explanation

What the study observed

What it does not establish

Open enough to check.

When OCR changes the script,
does it preserve the number?