Dataset Viewer
Auto-converted to Parquet Duplicate
copybooks
list
counts
dict
encoding_rules
list
generator
string
licence
string
[ { "copybook": "copybooks/c01_pk_basic.cpy", "id": "c01_pk_basic", "max_record_length": 10, "records": [ { "declared_codepage": "cp037", "encoded_with": "cp037", "fields": { "C01-NEG": { "expect": "NUMBER", "hazard": "comp3", ...
{ "copybooks": 41, "records": 49, "scored_fields": 166 }
[ { "field_type": "PIC X(n) DISPLAY", "rule": "n bytes, one EBCDIC code point per character, in the code page of the record.", "source": "LR, PICTURE clause (symbol X) and USAGE DISPLAY; IBM CDRA code page charts CCSID 37/500/1140" }, { "field_type": "PIC 9(n) / S9(n) DISPLAY (zoned decimal)", ...
gen_corpus.py
CC0-1.0

COBOL copybook decoding test corpus

Wrapping a COBOL system behind an API means some decoder turns fixed-width EBCDIC records into JSON. The risk is a value that looks plausible and is wrong: a negative amount that comes back positive, a packed decimal read one digit short, a code page that swaps a character. This corpus lets anyone check a decoder against records whose correct values are known by construction, not by trusting another decoder.

  • 41 copybooks, 49 records, 166 scored fields, including 8 must-fail control fields that a correct decoder should reject or flag instead of emitting a number.
  • Every record was encoded by gen_corpus.py with its own hand-written encoder (code-page tables, nibble, zone and binary packing). No decoder under test and no EBCDIC codec produced the bytes; Python's codecs only cross-check the hand tables, and the build aborts if they disagree.
  • The expected value of every field is the value that was encoded. Each encoding rule cites the IBM Enterprise COBOL Language Reference 6.4, the Programming Guide 6.4 or the z/Architecture Principles of Operation (see corpus/expected.json → encoding_rules).

Hazards covered

Hazard Fields
OCCURS DEPENDING ON (variably located items) 30
Baseline (plain DISPLAY and binary) 23
Zoned decimal with overpunched sign, SIGN LEADING / TRAILING / SEPARATE 17
REDEFINES 17
COMP-3 packed decimal 15
Implied decimal point (V) and scaling (P) 12
COMP / COMP-4 / BINARY 12
Code pages (CCSID 37, 500, 1140) 10
Alternate sign nibbles (A–F) and negative zero 9
COMP-5 values beyond the PICTURE digits 9
Must-fail controls (bad packed nibble, bad zone, wrong code page) and their valid siblings 12

Not covered yet: COMP-1/COMP-2 floating point, SYNC alignment, national/DBCS data, and RDW record framing. COMP values beyond the PICTURE digits are left out of COMP (they depend on the TRUNC compiler option) and tested under COMP-5, where the meaning is fixed.

Use it

python3 gen_corpus.py            # regenerate every copybook, record and expected.json byte for byte
# decode each record with your decoder, one JSON line per record:
#   {"record": "c01_pk_basic_r01", "fields": {"C01-NEG": "-12345", ...}}
#   {"record": "c41_mf_ascii_r01", "rejected": true}
python3 check.py corpus/expected.json your_output.jsonl

check.py uses the standard library only. It prints one verdict per field (MATCH, MISMATCH with both values, REJECTED) and a per-hazard tally. Numbers compare as exact decimals; text compares after removing trailing spaces.

Results against open-source decoders

We have run this corpus against seven configurations of open-source decoders. Those results are held until each project's maintainers have seen them. They will be published here afterwards as a dated record, together with any correction or context the maintainers send.

About

Published 28 September 2026 by CSOAI Ltd (Council of AI, councilof.ai), which measures whether published claims can be checked by a third party. This is a measurement resource: it does not certify, rate or rank any decoder, and it does not convert or remediate COBOL. Corrections: nicholas@csoai.org.

Licence: CC0 1.0. Use it for anything.

Objections, contact and corrections

To object to a row, ask for a re-check, request a correction or ask for a record to be withdrawn, email nicholas@csoai.org or use the appeals and dispute route at https://councilof.ai/dispute/. Corrections are listed in the corrections ledger at https://councilof.ai/api/corrections, with what changed and when. CSOAI Ltd (company no. 16939677, England and Wales) is the accountable publisher.

How to cite

CSOAI Ltd (Council of AI). COBOL copybook decoding test corpus. 2026. Hugging Face dataset csoai/cobol-copybook-decoding-corpus. https://huggingface.co/datasets/csoai/cobol-copybook-decoding-corpus

@misc{csoai_cobol_copybook_decoding_corpus,
  title        = {COBOL copybook decoding test corpus},
  author       = {{CSOAI Ltd}},
  year         = {2026},
  howpublished = {Hugging Face dataset, https://huggingface.co/datasets/csoai/cobol-copybook-decoding-corpus},
  note         = {Corrections: https://councilof.ai/api/corrections}
}

Licence: CC0-1.0. Attribution is appreciated, not required.

Corrections and verification

Downloads last month
130