Improving the way we build.

That ambition needs careful evaluation. A changed prompt, a new tool or a successful task is not enough on its own to show that a system has become more capable.

A development direction

We want improvements to be judged against a clear purpose, checked with appropriate evidence and retained only when the result justifies the change. The scope of a change, the conditions of the test and the limits of the evidence all matter.

Some Obinka products are intended to support this work. Others are applications through which we can explore how better development methods translate into useful software.

Questions we keep asking

• What changed? A prompt, a tool, a harness and a model's weights are different objects of change, and each needs its own description.
• How was it checked? The comparison, the conditions and the uncertainty belong with the result, not after it.
• Does it hold beyond the task? Gains selected on one set of work need testing on work that did not guide the change.
• What did it cost? Time, money and human review are part of whether a change was worth making.

Limits of this account

This page describes an intended approach. It is not a claim that every control listed here is already operational, and it is not a claim that any Obinka product improves itself in use without human review.