Under nondisclosure through Young Tech Systems. The client, the application and the domain it serves stay out; the method, the verification standard and the finding are what this entry can show, and they are the transferable part.
A specialized industrial engineering application, unsupported and unchanged for years. The analytical engine was still correct. What had gone stale was the reference data it reads, measured against a world that has been surveyed far better since that data was compiled.
The choice at that point is usually framed as replace or retire. There is a third option when the application’s reads can be separated from its writes: leave the engine alone and rebuild what it consumes.
Establishing that the seam exists
Separability is a claim about the software, and it had to be demonstrated rather than assumed. Every reference file the application consumed was inventoried, and each one established by evidence as read-only to the program. That distinction decides whether the approach is possible at all — a file the application also writes is not reference data, it is state, and swapping it underneath a running system corrupts it rather than updating it.
Round trip as the acceptance test
The formats were undocumented and had to be reverse engineered, but parsing was not the hard part. Verification was.
A converter whose output the application opens without complaint has proved very little. The dangerous failure is not a file that fails to load; it is a file that loads, runs, and returns an answer that is subtly wrong — in an engineering tool, silently and with full confidence. So the bar was set at an exact byte-for-byte round trip: decode an original file, re-encode it, and require the result to be identical to the bytes we started from. A converter that can reproduce the original exactly has understood the format rather than approximated it, and only then is a rebuilt file entitled to the trust the application already extends to the ones it shipped with.
Delivered across an air gap
The machine that runs the application is not on a network and should not be. Delivery was on read-only media, which changes what “finished” means: the artifact has to be complete and correct when it leaves, because there is no channel for a patch afterward and no telemetry coming back to say whether it worked.
Run as a reviewed pairing, not an automation exercise
The work was done as an expert-and-AI pairing under a written review discipline, and the discipline is the part worth recording. Acceptance criteria were fixed before anything was generated, so the test could not be quietly adjusted to fit what came out. Every artifact was verified against the original by round trip rather than by inspection, because inspection is exactly the check that scales worst and feels best.
The method was published along with seven cataloged failure modes and five preconditions, so the approach can be judged and reused rather than taken on trust.
The finding that generalizes
The binding constraint is expert review capacity, not generation capacity.
Generation is cheap and getting cheaper; the ability to establish that a given artifact is correct is neither. Any program that budgets these the other way round — sizing the work by how fast output can be produced — will produce volume in excess of what it can defend, which in an engineering context is worse than producing nothing.
The same method, run on other applications
The engagement itself is under nondisclosure and cannot be shown. What can be shown is the method applied elsewhere, against applications and projects belonging to this record rather than to a client.
Each view below is a propagation model whose engine was left exactly as it was, while the reference data underneath it — terrain, surface clutter, building footprints and canopy heights — was rebuilt from surveys that did not exist when the original data was compiled. The Kanawha valley view carries both states as switchable layers, so the original result and the rebuilt one can be set against each other directly.
What the approach buys.
- The engine keeps its standing. An analytical model that has been trusted for years does not have to be revalidated, because not a line of it changes. Replacing the application would put every past result back in question.
- The back catalog stays readable. Decades of accumulated project files still load, which is usually the real cost of migration and the one least often counted.
- The accuracy gain is the survey gap. The improvement is the distance between the data the application shipped with and what has been surveyed since — here, coarse legacy terrain against a 2018 LiDAR flight at two-metre grid. That margin is frequently larger than anything a newer engine would offer.
- It is reversible. The original reference data is kept and can be put back, so the decision is not one-way. Replacement is.
- It is checkable. Every rebuilt result can be run against the original on the same inputs, which is what makes the improvement arguable rather than asserted.
Outcome
Current reference data under an unchanged engine, delivered across an air gap, with the method published so it can be judged rather than taken on trust.
Interactive 3D signal views
Modeled coverage drawn over LiDAR terrain, roof shapes and tree crowns, with the EDX SignalPro result on the ground. Needs a browser with WebGL; best on a desktop.