Listing Bench 01: How to Transcribe a Public-Domain BASIC Type-In with OCR and Vision Models

Why BASIC type-in listings are difficult to digitize

A magazine type-in listing looks like ordinary printed text, but it behaves more like a compact technical specification. A misplaced zero, a substituted letter, or an omitted punctuation mark can change what a program does. The problem is especially noticeable in older BASIC listings, where variable names, abbreviated commands, line numbers, and compact expressions occupy narrow columns. A transcription that looks convincing on screen can contain several independent mistakes.

Original scans introduce additional difficulties. Paper aging can reduce contrast, ink can bleed through from the opposite page, and photocopying can turn small characters into dark blocks. Two characters that are visually distinct in a clean modern font may become almost indistinguishable in a deteriorated magazine scan. A listing's physical layout also matters: wrapped lines, continuation markers, and multiple columns can confuse software that expects ordinary paragraphs.

• Line numbers may be mistaken for program content.
• A zero and a capital O can become indistinguishable.
• Commas, periods, colons, and semicolons are easily confused.
• Quotation marks and apostrophes can disappear into scanning noise.
• A wrapped statement can be mistaken for two separate lines.

Prepare the source image before transcription

Start with the best available source. For a public-domain listing or material personally owned by the reader, a high-resolution scan is preferable to a photograph of a page displayed on another screen. A resolution around 300 to 600 pixels per inch is a useful starting range for printed material, with the higher end particularly helpful when the original characters are small. The appropriate setting depends on the paper, printing quality, and scanner.

Keep an untouched master image. Make working copies for rotation, cropping, contrast adjustment, and deskewing. Correct the page orientation before cropping the listing into individual columns. If the source has a gutter shadow, isolate that area carefully so image enhancement does not erase characters close to the binding. Avoid aggressive sharpening that creates artificial edges around punctuation.

Record the source page, magazine issue information when available, image dimensions, and any processing applied. These details help another reader locate the original material and distinguish an original character from an artifact introduced during image processing.

Choose OCR and vision models for different jobs

Traditional optical character recognition converts visible letter shapes into text. It can work efficiently on clean, consistently printed listings, particularly when the page is cropped into a single column and the text orientation is correct. Its weaknesses include unusual typefaces, low-contrast punctuation, and character combinations that are uncommon in ordinary prose. Some OCR systems also interpret programming syntax through language dictionaries, which can encourage inappropriate corrections.

Vision-language models can interpret a page image while considering its layout and programming context. A carefully worded instruction can ask for exact transcription, preservation of line numbers, and no correction of apparent mistakes. This is useful when the page includes unusual formatting or annotations. However, contextual interpretation creates another risk: a model may replace an unfamiliar expression with a plausible BASIC statement that was never printed.

The two approaches therefore provide different kinds of evidence. OCR offers a relatively mechanical recognition path, while vision models can provide an independent interpretation of difficult regions. Agreement between them is encouraging, but agreement alone does not establish that either reading matches the source.

• Use OCR for an initial machine-readable transcription.
• Use a vision model to interpret the original image independently.
• Request literal characters and original line ordering.
• Keep uncertain characters marked for human inspection.
• Never authorize automatic correction of programming syntax.

Compare the listing line by line

A useful comparison starts with line numbers, not with the apparent readability of the resulting program. Sort neither version automatically until the original sequence has been preserved. Compare the source image, OCR text, and vision transcription at the same line number. This makes it easier to detect a missing statement, a duplicated line, or an accidental change in program order.

For example, consider a printed statement containing 120 IF A=0 THEN 140. A transcription containing 120 IF A=O THEN 140 has only one incorrect character, but the variable reference may now mean something entirely different. Similarly, changing THEN to THFN can prevent a line from being interpreted as intended. These differences are easy to overlook in a long listing because the surrounding statements remain readable.

Classify discrepancies consistently. A misread character is different from a dropped line, and an invented line is different from a statement that has been split incorrectly. Maintain a separate record of unresolved regions. Do not silently repair the transcription to make the program appear more coherent. The objective is to reproduce the printed source faithfully, including apparent mistakes in the original.

• Misread character: a printed character becomes another character.
• Dropped line: an existing source line is absent.
• Invented line: the transcription contains a statement absent from the source.
• Structural error: line order or statement boundaries have changed.

Use checksums as a separate verification layer

Some magazine type-ins include a checksum, line total, or other numeric verification value. These features can provide a valuable independent check, but their interpretation depends on the original instructions. A checksum might be calculated from the characters in each BASIC line, the numeric line number, the ASCII values of characters, or a particular subset of the listing. A reader needs the documented algorithm before treating a matching value as meaningful.

For example, suppose an original instruction specifies a checksum based on the character values in each line, excluding line numbers and line-ending characters. Including the line number in the calculation would produce a different result even when the visible program text is correct. Differences in character encoding, spaces, and punctuation can also change the result. The expected checksum must therefore be associated with the exact calculation method and input representation.

A matching checksum is strong evidence against certain transcription mistakes, but it does not prove that every line was reproduced correctly. Different inputs can occasionally produce the same checksum. It also cannot independently establish that a magazine printed the intended program. Treat the checksum as one layer alongside direct comparison with the scan and, where appropriate, a known-good published listing.

Keep a reproducible transcription record

A useful archival record contains the original image, an unmodified OCR transcription, the independent vision transcription, and a corrected working copy. Record the software or service used, relevant model identification where available, image preparation settings, and the exact transcription instructions. Preserve the character encoding and line-ending convention as well. These details matter when a listing is moved between operating systems or imported into an emulator.

The final record should distinguish source text from editorial corrections. If the original page contains an apparent printing error, retain it in the diplomatic transcription and describe any proposed correction separately. A reader who wants to reproduce the magazine program should be able to see exactly what the magazine printed and which changes are interpretations.

For the next type-in, select one short column from a public-domain listing, create a clean working scan, and produce two independent transcriptions. Compare every line number and punctuation mark before calculating any documented checksum. Save the original image and all three text versions together.

Questions readers ask