Concrete Specifications

One of the most common—and costly—failure patterns in modernization is trying to rebuild a system and incorporate new requirements at the same time. It can create a cascade of unintended complexity.

‍

That’s why successful modernization starts with replication. It’s far safer and more likely to succeed if you first reproduce the existing system accurately before layering on new architecture patterns or innovation requirements. 

‍

In this post, we’ll explore how AI can accelerate that replication step: more importantly, how to wrangle a powerful, non-deterministic technology into producing a trustworthy behavioral copy.

‍

How can you trust AI to do what’s right and what works? Spoiler alert: you don’t.

‍

‍

Closed-Loop Validation

Mechanical Orchard’s mainframe modernization platform, Imogen, uses a closed-loop validation system that refines AI-generated code until it meets strict behavioral equivalence standards relative to the original application.

‍

Step 1: Real-time data capture

We integrate Imogen with system schedulers or use real-time monitoring to capture inputs and outputs from production workloads in the existing system. These can include:

‍

• relational database data (with or without change data capture)
• datasets / files (including generational datasets)
• MQ messages
• job parameters and other relevant runtime configuration

‍

Once captured, the data is stored in a way that allows all relevant data for a given execution of a workload to be 'replayed' against the modern system: we call these replica packets.

‍

‍

Step 2: Contextualization and code generation

Context is critical when it comes to recreating the existing behavior. It can include, among other things:

‍

• the relevant source code for the workload, or a portion of it
• job execution parameters and metadata
• information about how to replicate legacy behaviors in the new system and framework, such as how to send emails, write MQ messages, or make database updates

‍

Imogen’s AI engine uses this context to characterize the existing system, which then guides the code generation process. 

‍

‍

Step 3: Convergence

The resulting code is automatically integrated with the Imogen runtime framework and then executed against the replica packets captured earlier. Imogen runs the generated code against the data packets, captures the results, including compiler messages, runtime errors, and information about whether the outputs and side effects matched those described in the data packets. 

‍

If the resulting code fails to match the original workload in terms of equivalence, Imogen automatically tries again (and again if needed), each time including information about how the outputs were not equivalent, until the process succeeds, as depicted by all replica packets ‘going green’.

‍

‍

Step 4: Review and Equivalence

The converted workload is now ready for review by developers, who can then promote the code into the equivalence environment. In this environment, the new workload is run alongside the legacy one, where it is given the same inputs, and outputs are again checked for equivalence. 

‍

The new workload is run in equivalence mode until the code meets performance, non-functional, and behavioral requirements, at which point it can go into production and the corresponding legacy workload decommissioned. 

‍

‍

‍

Verify, then trust

Mechanical Orchard’s closed-loop validation model makes sure that any AI-generated code is not just theoretically correct but rigorously validated against real-world data within actual workflows and business processes. The code Imogen generates is clean and idiomatic, making it easy to read, understand, and maintain for human or AI developers.

‍