The Human Is The Override

Four AI Reviews Found the Bugs. A Pilot Found the Fix.

TL;DR

Sixteen days, 71 build passes and 91 simulator sessions produced more than a formation-flying trainer. The project became a live experiment in human-AI collaboration. AI reviews uncovered hidden measurements and flawed assumptions, but the breakthrough that mattered most came from a simple pilot question: are we solving this the way a human actually would?

  • Adversarial review was the highest-value use of AI in the project.
  • Multiple AI reviews found important problems and hidden assumptions.
  • The most important breakthrough came from operator experience.
  • The project reproduced organisational scaling problems at team size two.
  • Operational memory mattered more than model intelligence.
  • Hybrid teams outperform humans or AI operating alone.

A live laboratory for human-AI collaboration

Telemetry Tracking

The initial aim was straightforward: create a prototype AI-led aircraft for formation practice in Microsoft Flight Simulator. The project grew into experimenting with control laws, developing flight-test workflows, analyzing telemetry, and building a scoring system for training runs. What began as a software prototype turned into an engineering process that highlighted the importance and benefits of hybrid work.

When Reviews Became the Bottleneck

Repeated failures triggered a formal review process. Every review received a cold brief, evidence recordings, design status and a focused set of questions. Predictions were written before tests. Claims were checked against evidence. The goal was not agreement. The goal was finding what everyone was missing.

Two Problems the Models Found

The first breakthrough came from a measurement hidden in plain sight. Once identified, a long-standing error collapsed from tens of feet to less than one. The second came from challenging a textbook assumption embedded in a control law. Both fixes worked immediately because they revealed mechanisms rather than guesses.

The Bug Nobody Could See

One aircraft repeatedly failed. Every review treated the problem as mathematics. Every proposed solution became more sophisticated. None worked. The breakthrough came from stepping back and asking how a pilot would handle the situation. The system was fighting the aircraft instead of cooperating with it. The next flight completed successfully. That was the moment the project clarified its real lesson: better reasoning starts with the right frame.

The Human Is The Override

Human is the override

Every model optimised inside the frame it had been given. The frame itself was wrong. Domain expertise, operational experience, and understanding intent supplied the missing context. The human contribution was not generating answers. The human contribution was deciding which question actually mattered.

The Invoices

The economics were as interesting as the engineering. Sixteen days of daily development cost roughly the same as a single frontier-model review. The expensive review delivered valuable findings. It still missed the decisive fix. Cost and intelligence are not the same thing.

The Research Says the Same Thing

Industry research increasingly shows that AI accelerates code production faster than organisations can validate it. Output rises. Review load rises with it. The bottleneck moves downstream. This small project mirrored the same pattern at miniature scale.

A Team of Two Hit Every Scaling Problem

Despite consisting of one human and several AI systems, the project encountered onboarding issues, process drift, trust debt, documentation failures and coordination overhead. The same organisational problems that appear in larger teams emerged almost immediately.

Five Practices That Fixed It

The most useful improvements were simple:

  • Session cards
  • Evidence-backed handovers
  • Single ownership of important information
  • Bounded rework cycles and
  • Filing evidence immediately.

Together they created operational memory that neither humans nor models could reliably maintain alone.

Operational Memory Is The Key

The reviews only worked because operational memory existed. Every reviewer inherited logs, evidence, decisions, and context. Without that memory layer, review quality would have collapsed. The project became a demonstration of why operational knowledge needs structure. In a wider business context, this is why applications like MimeOps exists.

What Comes Next

Hybrid Formation Teamwork

Hybrid Formation Teamwork

The future is unlikely to belong to a single model tier. Practical teams will combine inexpensive workhorse models with targeted use of frontier systems at critical decision points. The skill is not choosing the smartest model. The skill is knowing when a different perspective is worth buying.

Final Thought

The question was never whether AI was useful. It clearly was. The question was whether AI could replace the person responsible for deciding what mattered. After sixteen days, ninety-one flight sessions and four independent reviews, the answer remained remarkably clear: the human is still the override.

Written By

Pin It on Pinterest