HC3 modernization paper
Why generic agentic AI fails at mainframe scale.
LLMs can explain code, but they should not be asked to discover an enterprise system from fragmented context. At mainframe scale, deterministic system mapping must come before LLM reasoning.
Prepared for discussion and validation. Quantitative claims are limited to provided case facts; hallucination reduction should be measured during validation before being stated as a percentage.
01 · Situation
Useful on small files. Unreliable at estate scale.
In 2026, a large bank launched a modernization initiative across 14,155 mission-critical mainframe applications and initially evaluated Claude and Microsoft Copilot to accelerate code understanding, documentation, onboarding, and planning. The tools performed well on small files and isolated repositories where the required context fit within the model’s working window, but their reliability declined as the scope expanded across interconnected enterprise systems.
02 · Problem
The IDMS stress test.
IDMS exposed the weakness of LLM-only modernization because it overlaps with COBOL and Assembler while introducing its own syntax, runtime behavior, dialogs, and data access patterns. The team needed deterministic handling, not AI interpretation alone.
Where trust breaks
From mainframe estate to a trust and risk problem.
- Mainframe estate
COBOL, HLASM, PL/I, IDMS, CICS, BMS, ADSO, DB2, Java
- Generic AI
Works on isolated code samples
- Cross-program analysis needed
Transaction flows, shared dependencies
- Hallucinations
Outputs not traceable to code
- Trust & risk problem
Must be accurate, auditable
03 · Experiment
Map first. Reason second.
In June 2026, the bank began validating Adapts HC3 across the full mainframe estate — COBOL, HLASM, PL/I, IDMS, CICS macros, BMS maps, ADSO dialogs, database calls, and cross-repository dependencies — to test whether a deterministic map-then-reason approach could solve the traceability problem conventional LLM-led analysis could not.
| Validation area | Bank question | Grounding needed |
|---|---|---|
| Parsing & coverage | Is there native parsing logic for IDMS, CICS macros, BMS maps, ADSO dialogs, and DB2 call syntax — or does it rely on AI to interpret them? Are ADSO sources PDS text members or compiled-only in the CA dictionary? | Per-construct evidence: parser/extractor logic, supported constructs and limitations, fallback path, confirmed source format, extraction method, and validation sample. |
| Unified queryable estate | Can COBOL, HLASM, PL/I, and IDMS be processed together, with users able to query across multiple systems and application boundaries? | Single graph-backed intelligence layer combining open-source COBOL/PL/I samples with bank-provided IDMS dummy code, spanning systems and applications. |
| Impact analysis | Can a field or variable change be traced across dependencies? | Dependency traversal showing upstream/downstream programs, data stores, screens, and calls. |
04 · Deterministic approach
Build solid ground before the model speaks.
Same divergence point both times: LLM-only starts fine but has nothing solid underneath it once scope widens, while Adapts builds the solid ground first — which is what enabled the >95% hallucination reduction observed in validation.
Hallucination reduction should be measured during validation before being stated as a percentage outside this discussion context.
| Dimension | LLM-only risk | HC3 grounding |
|---|---|---|
| Discovery method | Prompt-driven inference over partial files or repository context. | Static analysis first; graph construction before LLM reasoning. |
| Context boundary | Partial files and repo windows that break across systems. | Connected system map spanning repositories, languages, and dependencies. |
| IDMS handling | Risk of treating IDMS syntax as natural-language or COBOL-like context. | Deterministic parsing/extraction for IDMS constructs, then LLM explanation on verified facts. |
| Output quality risk | Plausible explanations may not be traceable to source relationships. | Insights linked to programs, variables, calls, maps, and graph edges. |
| User trust model | User must manually validate model assumptions. | User validates evidence-backed dependencies and impact paths. |
The pattern
What fails. What works.
LLM-only
- Prompt-driven inference
Partial files, repo context
- Context degrades
Fails to scale across systems
- Hallucination risk
Explanations not traceable
Adapts HC3
- Static analysis first
Graph built before LLM
- Connected system map
Repos, languages, dependencies
- Evidence-based reasoning
>95% fewer hallucinations
05 · Core lesson
Enterprise systems need a verified map.
The bank experience showed that agentic modernization does not fail because LLMs cannot understand code. It fails when LLMs are asked to understand enterprise systems without a verified system map.
For small files, LLM-only modernization can be useful. For isolated repositories, AI can accelerate documentation. But for enterprise mainframe modernization across thousands of applications, multiple languages, IDMS, CICS, BMS, ADSO, database calls, and cross-repository dependencies, modernization requires deterministic system intelligence.
See it on your codebase
See Adapts HC3 on a mainframe-like estate.
30-minute technical walkthrough with an enterprise architect. No slides · a live demo on real code.