Evaluation Kept Failing Before the First Test Ran.
Nugalaxy did not start as an evaluation practice. It started with AI systems that had to work, and with the ordinary problem of showing that they did.
The failures were never where we went looking for them. The scoring was fine. The tooling was fine. What was missing sat further upstream: nobody had written down what correct behavior actually meant. So the evaluation measured whatever was easy to measure, the result came back green, and the requirement that mattered was never turned into a test at all.
What the system was actually supposed to do, never written down.
That gap is the whole of what this practice works on. We define what correct means first, then build the criteria, cases, harness, and evidence that make an evaluation result mean something to the person who has to sign off on it.
What We Actually Believe.
A score is not evidence.
A number tells you that something happened. Evidence tells you what was required, what was tested, what the system did, and what that result demonstrates to someone who did not build it.
The hard part is upstream.
Deciding what has to be proven is harder than running the evaluation, and it is the part that gets skipped. Tooling does not rescue an evaluation that was aimed at the wrong thing.
Evaluation is engineering.
It needs a harness, a repeatable environment, and results you can re-run after a prompt, model, or tool change. A spreadsheet and a review meeting are not an evaluation.
Show the work.
We publish the methodology, the guides, and the library instead of describing them. You can read how we think before you decide whether to talk to us.
Small, On Purpose.
Nugalaxy is small, and the person writing the guides is the person doing the engineering. That is deliberate. The methodology stays specific because it comes out of real systems rather than a content calendar, and the work we take on feeds it directly. Nothing here is generalized from someone else's whitepaper.
It is also why harness engineering is written out in public rather than kept as a house method. A discipline nobody can read is not a discipline.
Computed, Not Drawn.
The same rule runs all the way down to the logo. Every planet in this universe is generated from math, not illustrated by hand. The cyan caltrop you see is a shape our own code produces. Form follows the engineering, with nothing drawn on top to hide it, which is the same reason we hand over evidence instead of assurances.