Back to technical studies
TECHNICAL STUDY

Back Engineering: A ZIP Is Not an Explanation

Working backwards through an AI package to separate real evidence from a plausible-sounding diagnosis.

Sometimes the best way to understand an AI system is to stop reading its promises and examine the package. In February 2026, I used an Agency MyGPT package as a Back Engineering experiment.

Four useful questions

What was the system intended to do? Which files actually controlled its behavior? Where did its rules, specialist roles and tools meet? And what would it take to rebuild it without reproducing the same mistakes?

That approach led toward inventories, authority maps, failure cases and reproducible reviews. But the archive records a telling limitation. In one session, the assistant could see the ZIP filename and methodological templates without being able to inspect the ZIP's actual folder tree.

An honest stop is a useful result

At that point, a method could be proposed but the hidden package architecture could not be certified. No inspected file means no verified statement about its contents. No executed test means no reliable performance claim.

The practical benefit wasn't a magical reconstruction button. It was a review discipline: identify evidence first, separate design from behavior, keep gaps visible and make the next revision inspectable.

A system that can withstand scrutiny matters more than one that only sounds convincing.

The historical record supports the analysis approach and documented package iterations; not a finished, independently benchmarked back-engineering product.

All writing