On evaluation

The AI industry is grading its own exam.

Every few weeks a lab announces a new state-of-the-art on a benchmark it optimized toward, scored by a rubric it helped write. That's teaching to the test. It tells you a model can pass the exam. It tells you nothing about whether it does your job. We measure the other thing — with reinforcement baked into the controllers that run the work and the design systems that feed the knowledge graph.

Teaching to the test

Four ways a benchmark lies.

The measure becomes the target

Goodhart's law, in production: once a score is the goal, it stops measuring the thing it stood for. Models are tuned until the number goes up, not until the work gets better.

Contamination

Public test sets leak into training data. A model that has seen the answers isn't reasoning; it's remembering. You can't tell which from a leaderboard.

Self-reported, self-refereed

The vendor picks the benchmarks, the prompts, and the harness, then reports the wins. The exam and the examiner share a payroll.

It isn't your distribution

A high score on a generic exam says nothing about your documents, your rules, your edge cases. The only test that matters is the one made of your real work — and no lab has that.

Our answer

Graded by the domain, not the leaderboard.

We don't chase a public score. We build the reward signal into the machinery itself, so the system is judged — continuously — by whether real work in your domain gets accepted.

RL in the controllers

Every task the actor system runs posts a quality signal back. The router scores providers and packs on a running mean per capability and sends the next job to whatever is actually performing — here, now, on this kind of work. A rejected output isn't just a loss; it's a policy update the next attempt inherits. The reward is real acceptance, not a held-out set.

RL in the design systems & knowledge graph

Every accepted result becomes a grain in the knowledge graph, and the design systems that shape the work carry the domain's rules forward. Domain quality scores recalibrate from real performance, so the system gets more selective and more accurate exactly where it has evidence. The graph is the exam, and it's written in your outcomes.

The loop that grades itself honestly

01

Real work runs through the controllers.

02

A quality gate scores the output before it ships.

03

Accept or reject becomes reward — a grain in the graph.

04

Domain scores and routing recalibrate.

05

The next task is graded by everything the last one learned.

Evals, reward models, audit logs, and rollout gates aren't a launch-week slide — they're the governance layer of the stack, running on every request. And because the deterministic core is byte-reproducible, the surface that even needs judging is small: we grade the residual, not the whole app. Where this sits in the stack →


A leaderboard win is a photo of a model on its best day. We ship the film of it doing your job.

Bring your real work. We'll grade against that.

Request access How Mempacks™ stay current