Research

Claims about AI-driven productivity in software delivery are easy to make and hard to check. Most of them arrive as best-case anecdote, with the methodology either absent or reconstructed after the result was already known.

This section is an attempt to do the opposite: fix the hypothesis, the problem-selection criteria, and the measurement boundary in public before any run happens, then report whatever comes back — including a null result.

How this is organized

The protocol is the single source of truth for methodology. It defines the hypothesis, how a problem gets selected, what the three arms are, where the clock starts and stops, and what would count as disconfirmation. It carries its own changelog, so any post-hoc revision is dated and visible rather than silently applied.

Each scenario is one problem run against that protocol. Scenario pages hold only what is specific to that run — the problem, why it was selected, the resulting backlog, and the measured outcome. They do not restate the methodology; they cite it. A scenario is published as soon as its problem is selected, while its results are still empty, so that the record shows the commitment was made before the data existed.