Better, or just different?
Compare two versions and see which changes are real and which are noise.
Tesryn Lab · AI evaluation
Research Benchmark
Know whether a change to an AI research assistant actually made it better.
In use at Tesryn First built for teams improving AI research and search tools · 2026
01 — What it is
A fixed, repeatable test for AI assistants that read documents and answer questions. It scores answers, catches the known ways they go wrong, keeps a sealed set nobody tunes against, and compares any two versions side by side. It was built before the assistant it measures, so the assistant can't be shaped around its own scoreboard.
The moment it earns its keepA change that felt like an improvement turns out to be noise, and you find out in an afternoon instead of after a quarter.
What it proves we can do
02 — What it can do
Four things it does.
Knows how answers go wrong
Catches copying whole documents, confident guesses and missing evidence.
A sealed exam
A held-back set that no change is ever tuned against.
Regenerates exactly
The whole test rebuilds identically, or with fresh material of the same shape.
03 — How it connects
What it shares with the rest of the Lab.
The same things we know how to do keep coming back in different industries. This is where this one sits: what it proves, and what else proves the same thing.
Knowing whether a change actually helped Conditions written before the run, a sealed test, and a verdict that stands either way. Also proved by Capability Distillation, Research Runtime.
Comparing only where evidence can tell things apart Differences smaller than what the method can measure are reported as ties, not wins. Also proved by Magnetic Digital Twin, Capability Distillation.
Every claim traced to its source Nothing is asserted without the passage, part number or measurement behind it. Also proved by Podcast Maker, Engineering Library, Reconstruction Research Studio, Part Finder, Application Engineering, Research Runtime.
The whole map — six sections, and everything that proves them.
04 — What it's held to
- Reference systems, good and deliberately bad, show what a score means.
- Runs offline, with no outside service needed.
What it isn't. Not an assistant itself. It's the instrument you point at one.
05 — The recipe stays in the lab
We'll tell you what it can do. Not how it's built.
Yours would be your own build, on your systems, and you'd own it outright: code, documentation, no lock-in. What we bring to it is the practice — we've solved this shape of problem before, so you're not paying for the first attempt. The method stays ours. And if what you need isn't here at all, that's the usual case — it's how every skill on this page started.