From Black Field to Blueprint

Key Challenges

1. Lacking Supply Code: legacy understanding is already advanced if you
   have supply code and an SME (in some type) to place the whole lot collectively. When the
   supply code is lacking and there are not any consultants it’s a good better problem.
   What’s left are some compiled binaries. These will not be the latest binaries that
   are simple to decompile as a consequence of wealthy metadata (like .NET assemblies or JARs), these
   are even older binaries: the sort that you just would possibly see in previous home windows XP beneath
   C:Home windowssystem32. Even when the database is accessible, it doesn’t inform
   the entire story. Saved procedures and triggers encode many years of amassed
   enterprise guidelines. Schema displays compromises made primarily based on context unknown.
2. Outdated Infrastructure: OS and DB reached finish of life, gone its
   LTS. Software has been in a frozen state within the type of VM resulting in
   vital danger to not solely enterprise continuity, additionally considerably growing
   safety vulnerability, non compliance and danger legal responsibility.
3. Institutional Data Misplaced: whereas hundreds of finish customers are
   repeatedly utilizing the system, there may be hardly any enterprise data out there
   past the occasional assist actions. The reside system is the most effective supply of
   data. The one dependable view of performance is what customers see on display screen.
   However the UI captures solely the “final mile” of execution. Behind every display screen lies a
   tangled net of logic deeply built-in to a number of different core methods. It is a
   frequent problem, and this technique was no exception, having a historical past of a number of
   failed makes an attempt to modernize.

Our Objective

The target is to create a wealthy, complete useful specification
of the legacy system with no need its unique code, however with excessive
confidence. This specification then serves because the blueprint for constructing a
trendy alternative utility from a clear slate.

• Perceive total image of the system boundary and the mixing
  patterns
• Construct detailed understanding of every useful space
• Establish the frequent and distinctive eventualities

To make sense of a black-box system, we wanted a structured solution to pull
collectively fragments from completely different sources. Our precept was easy: don’t
attempt to get better the code — reconstruct the useful intent.

Our Multi Lens Method

It was a 3 tier structure – Internet Tier (ASP), App Tier (DLL) and
Persistence (SQL). This structure sample gave us a soar begin even with out
supply repo. We extracted ASP recordsdata and DB schema, saved procedures from the
manufacturing system. For App Tier we solely have the native binaries. With all
this info out there, we deliberate to create a semi-structured
description of utility conduct in pure language for the enterprise
customers to validate their understanding and expectations and use the validated
useful spec to do accelerated ahead engineering. For the semi-structured
description, our method had broadly two elements

1. Utilizing AI to attach dots throughout completely different knowledge sources
2. AI assisted binary Archaeology to uncover the hidden performance from
   the native DLL recordsdata

UI Layer Reconstruction

Looking the present reside utility and screenshots, we recognized the
UI components. Utilizing the ASP and JS content material the dynamic behaviour related
with the UI ingredient could possibly be added. This gave us a UI spec like beneath:

What we seemed for: validation guidelines, navigation paths, hidden fields. One
of the important thing challenges we confronted from the early stage was hallucination, each
step we added an in depth lineage to make sure that we cross examine and ensure. In
the above instance we had the lineage of the place it comes from. Following this
sample, for each key info we added the lineage together with the
context. Right here the LLM actually sped up the summarizing of enormous numbers of
display screen definitions and consolidating logic from ASP and JS sources with the
already recognized UI layouts and area descriptions that may in any other case take
weeks to create and consolidate.

Discovery with Change Knowledge Seize (CDC)

We deliberate to make use of Change Knowledge Seize (CDC) to hint how UI actions mapped
to database exercise, retrieving change logs from MCP servers to trace the
workflows. Setting constraints meant CDC might solely be enabled partially,
limiting the breadth of captured knowledge.

Different potential sources—reminiscent of front-end/back-end community site visitors,
filesystem modifications, further persistence layers, and even debugging
breakpoints—stay viable choices for finer-grained discovery. Even with
partial CDC, the insights proved beneficial in linking UI conduct to underlying
knowledge modifications and enriching the system blueprint.

Server Logic Inferance

We then added extra context by supplying
typelibs that have been extracted from the native binaries, and saved procedures,
and schema extracted from the database. At this level with details about
format, presentation logic, and DB modifications, the server logic may be inferred,
which saved procedures are seemingly referred to as, and which tables are concerned for
most strategies and interfaces outlined within the native binaries. This course of leads
to an Inferred Server Logic Spec. LLM helped in proposing seemingly relationships
between App tier code and procedures / tables, which we then validated by way of
noticed knowledge flows.

Prior Makes an attempt

Earlier than we arrived at this method, we tried

• brute-force technique: Added all meeting capabilities right into a workspace, and used
  the LLM agent to make it humanly readable pseudocode. Confronted a number of challenges
  with this. Ran out of the 1 million context window as LLM tried to finally
  load all capabilities as a consequence of dependencies (references it encountered e.g. perform
  calls, and different capabilities referencing present one)
• Break up the set of capabilities into a number of batches, a file every with 100s of
  capabilities, after which use LLM to research every batch in isolation. We confronted loads
  of hallucination points, and file measurement points whereas streaming to mannequin. A couple of
  capabilities have been transformed meaningfully however loads of different capabilities did not make
  any-sense in any respect, all seemed like related capabilities, on cross checking we
  realised the hallucination impact.
• The subsequent try was to transform the capabilities one by one, to
  guarantee LLM is supplied with a contemporary slim window of context to restrict
  hallucination. We confronted a number of challenges (API utilization restrict, charge
  limits) – We could not confirm what LLM translation of enterprise logic
  was proper or mistaken. Then we could not join the dots between these
  capabilities. Fascinating observe, we even discovered some C++ STDLIB capabilities
  like
  std::vector::insert
  on this method. We discovered loads have been truly unwind capabilities purely
  used to name destructors when exception occurs (stack
  unwinding)
  destructors, catch block capabilities. Clearly we wanted to deal with
  enterprise logic and ignore the compiled library capabilities, additionally combined
  into the binary

After these makes an attempt we determined to alter our method to slice the DLL primarily based
on useful space/workflow somewhat than contemplate the entire meeting code.

Discovering the related perform

The primary problem within the useful space / workflow method is to discover a
hyperlink or entry level among the many 1000s of capabilities.

One of many out there choices was to fastidiously take a look at the constants and
strings within the DLL. We used the historic context: late Nineteen Nineties or early 2000
frequent architectural sample adopted in that interval was to insert knowledge into
the DB: was to both “choose for insert” or “insert/replace dealt with by saved
process” or by way of ADO (which is an ORM). Apparently we discovered all of the
patterns in several elements of the system.

Our performance was about inserting or updating the DB on the finish of the
course of however we could not discover any insert or replace queries within the strings, no
saved process to carry out the operation both. For the performance we
have been searching for, it occurred to truly use a SELECT by way of SQL after which
up to date by way of ADO (activex knowledge object microsoft library).

We received our break primarily based on the desk title talked about within the
strings/constants, and this led to discovering the perform which is utilizing that
SQL assertion. Preliminary take a look at that perform did not reveal a lot, it could possibly be
in the identical useful space however a part of a special workflow.

Constructing the related subtree

ASM code, and our disassembly instrument, gave us the perform name reference
knowledge, utilizing it we walked up the tree, assuming the assertion execution is one
of the leaf capabilities, we navigated to the father or mother which referred to as this to
perceive its context. At every step we transformed ASM into pseudo code to
construct context.

Earlier once we transformed ASM to pseudocode utilizing brute-force we could not
cross confirm whether it is true. This time we’re higher ready as a result of we all know
to anticipate what could possibly be the potential issues that might occur earlier than a
sql execution. And use the context that we gathered from earlier steps.

We mapped out related capabilities utilizing this name tree navigation, generally
we’ve to keep away from mistaken paths. We discovered about context poisoning in a tough
manner, in-advertely we handed what we have been searching for into LLM. From that
second LLM began colouring its output focused in direction of what we have been wanting
for, main into mistaken paths and eroding belief. We needed to recreate a clear
room for AI to work in throughout this stage.

We received a excessive degree define of what the completely different capabilities have been, and what
they could possibly be doing. For a given work movement, we narrowed down from 4000+
capabilities to 40+ capabilities to cope with.

Multi-Move Enrichment

AI accelerated the meeting archaeology layer by layer, move by move: We
utilized multi move enrichment. In every move, we both navigated from leaf
node to prime of the tree or reverse, in every step we enriched the context of
the perform both utilizing its dad and mom context or its youngster context. This
helped us to alter the technical conversion of pseudocode right into a useful
specification. We adopted easy methods like asking LLM to present
significant technique names primarily based on identified context. After a number of passes we construct
out your complete useful context.

Validating the entry level

The final and demanding problem was to substantiate the entry perform. Typical
to C++, digital capabilities made it tougher to hyperlink entry capabilities in school
definition. Whereas performance seemed full beginning with the foundation node,
we weren’t certain if there may be some other further operation taking place in a
father or mother perform or a wrapper. Life would have been simpler if we had debugger
enabled, a easy break level and evaluation of the decision stack would have
confirmed it.

Nonetheless with triangulation methods, like:

1. Name stack evaluation
2. Validating argument signatures and the the return signature within the
   stack
3. Cross-checking with UI layer calls (e.g., associating technique signature
   with the “submit” name from Internet tier, checking parameter varieties and utilization, and
   validating towards that context)