Automated Program Logic Manuals (PLM) for legacy COBOL systems. Deterministic fact extraction meets constrained AI. Result: production-ready documentation with zero entity fabrications.
The average COBOL programmer is nearing retirement age. With 85% of universities having dropped COBOL from their curriculum, the pipeline of replacements has dried up. Organizations report that finding qualified mainframe talent is their top challenge.
Millions of lines of COBOL run banking, insurance, and government systems with little or no documentation. When the developers who understand a program retire, the knowledge goes with them, leaving maintenance teams guessing.
Generic AI documentation tools fabricate field names, invent paragraph references, and create plausible-sounding but wrong descriptions. In mission critical environments, fabricated documentation is worse than no documentation.
Three independent parser layers break each COBOL program into a Structured Technical Baseline (STB), a machine-readable JSON containing every verifiable fact in the source.
The AI reads source code independently and uses the STB as a cross-reference, grounding every statement in specific source lines. Archetype-aware specifications ensure each program type receives appropriate treatment.
Methodology, rubric, per-program scores and the machine-readable scores.csv are published in full, so the number can be recomputed rather than taken on trust. Read the benchmark overview, the full results or the methodology, or review the whole benchmark on GitHub.
| Archetype | Score |
|---|---|
| CICS Other | 99.4 |
| Batch File Maintenance | 99.1 |
| Batch Report | 99.1 |
| CICS Inquiry | 98.6 |
| CICS Menu | 98.4 |
| CICS Data Entry | 98.3 |
| Batch Utility | 97.7 |
| Called Subprogram | 97.7 |
A maintenance programmer receiving these PLMs can trust the field names, control flow descriptions, I/O operations, and business logic explanations without needing to cross-reference source code for basic comprehension.
Minor deductions below 100 typically reflect inferences about business context that cannot be mechanically verified from source code alone, such as who consumes a report or when a batch job is scheduled. These are appropriately qualified in the PLMs rather than stated as fact.
A score of 98.7 across 42 programs spanning 8 archetypes means the generated documentation is consistently accurate enough for production maintenance use.
Every fact traced to source lines. Every entity evaluated against the Structured Technical Baseline. Both the deterministic scaffold and AI narrative are delivered, giving compliance teams a verifiable evidence trail.
Every named reference — field names, paragraph names, CICS commands, file names, dataset names — is cross-referenced. An output containing a name that doesn't appear in the source is flagged before delivery. The PLM is rejected, not patched. There are no fabricated facts.
Batch reports, CICS data entry, file maintenance, called subprograms; each program type receives specialized treatment with rubrics tailored to its maintenance patterns.
When onboarding a new repository, a gap detector identifies COBOL language patterns the parsers haven't yet encountered, ensuring extraction quality is evaluated before documentation is delivered.
Clients receive both the deterministic PLM (verified factual scaffold with zero AI) and the AI-synthesized PLM (narrative documentation with business context). A verified baseline alongside readable documentation.
Source expansion, deterministic parsing, context pack generation, constrained AI synthesis, and automated evaluation. Our entire documentation pipeline runs without manual intervention.
Give regulators and auditors verifiable, line-referenced documentation for critical financial systems. Every fact is traceable to source.
Cut onboarding time for new developers. PLMs explain business logic, control flow, and data structures without requiring source code archaeology.
Comprehensive program documentation is the prerequisite for any migration initiative whether refactoring to microservices or converting to modern languages.
An annotated walkthrough of a real PLM section by section — the program overview, every file touched, control flow, status checks and failure paths, copybooks, and how it runs in production. Callouts explain what each block means.
Open annotated PLM → PIPELINE WALKTHROUGHA scroll-through tour showing how deterministic parsing builds the Structured Technical Baseline, and how constrained AI synthesis turns that base into the readable Program Logic Manual — with zero entity fabrications.
Open walkthrough → FULL PDF · CBTRN03CThe full production-ready Program Logic Manual for CBTRN03C, the daily transaction posting batch program from AWS CardDemo. Same format your engineers and auditors would receive on day one of an engagement.
Download PDF →Every program, call and control transfer, every dataset and DB2 table. Hover any program for a plain-English summary, then click to drill from the system graph into a program's job & file flow and down to the data flow between its paragraphs. 83 linked views.
Open the flow map → WHO TOUCHES WHAT34 programs against 38 data stores. Every Create, Read, Update and Delete marker opens the paragraph and source-line citations that prove that access — the evidence an auditor asks for, not an assertion.
Open the matrix → BLAST RADIUSPick a program or data store and trace everything it reaches upstream and downstream, with a citation on every hop. Unresolved dynamic targets are excluded rather than guessed. Plus a static findings report covering what never runs.
Open impact analysis →