Mezure

Evaluate competence on the agentic era.

Evaluated lines of code:Competence percentage:
Book a demo

AGI research agent eval

AGI research agent eval
32:14

I ran the latest trace and the harness is on the uncertainty branch.

Your candidate

the abstention test passes now, but the agent still sounds certain when the docs are thin.

inspect sparse-doc trace

5173/evals/research-agent

Training loss

AGI research agent harness

0.18

loss

0.18inittuneconverge

The full workspace, already instrumented.

We place your candidates in a controlled and audited browser-based workspace with your chosen starting codebase, model, instructions, and a terminal. Assessments are fully laid out for your team to judge, with our optional insights. Requires no setup on the candidate's end, just a link.

Define competence on your terms

Answer a few questions and Mezure assembles the assignment, environment, and rubric for you. Tune anything you like, then send candidates a single link. The bar is yours to set.

Create assessment

Autosaved to your workspace. Close this dialog and continue from the sidebar.

What kind of engineering competence are you trying to evaluate?

Start from a common assessment shape. You can still edit every answer as the flow continues.

Debugging
Feature build
From scratch

Measure competence on the correct abstraction layer.

As intelligence commoditizes, engineering keeps sliding further toward research across software and hardware. Models are now capable enough to raise the baseline, demanding a higher and more abstract competence judgment layer. Today every company is trying its best at a blind shot; there is no golden standard for this. Mezure is an attempt at one.

Candidate review

Gregory House

All changes from the candidate

Search files

src/search/rank-results.ts+41 -8
- const freshnessBoost = result.cached ? 0.18 : 0;
+ const freshnessBoost = result.cached ? 0.04 : 0.16;
const semanticScore = cosineSimilarity(query, result.embedding);
+ const keywordScore = exactMatchBoost(query, result.title);
return semanticScore + freshnessBoost + keywordScore;
M: Off-by-one risk if the ranked list is empty.
M: Ask why keyword boost is safer than another semantic cutoff.
src/search/cache.ts+27 -6
- const freshnessBoost = result.cached ? 0.18 : 0;
+ const freshnessBoost = result.cached ? 0.04 : 0.16;
const semanticScore = cosineSimilarity(query, result.embedding);
+ const keywordScore = exactMatchBoost(query, result.title);
return semanticScore + freshnessBoost + keywordScore;
M: Candidate changed cache behavior without documenting rollout risk.
src/search/scoring.ts+18 -4
- const freshnessBoost = result.cached ? 0.18 : 0;
+ const freshnessBoost = result.cached ? 0.04 : 0.16;
const semanticScore = cosineSimilarity(query, result.embedding);
+ const keywordScore = exactMatchBoost(query, result.title);
return semanticScore + freshnessBoost + keywordScore;
tests/search-ranking.test.ts+32 -4
describe("ranking regression", () => {
+ it("prefers fresh exact matches over stale semantic hits", () => {
+ expect(rankResults(query, fixtures)[0].id).toBe("fresh");
docs/ranking-notes.md+14
- const freshnessBoost = result.cached ? 0.18 : 0;
+ const freshnessBoost = result.cached ? 0.04 : 0.16;
const semanticScore = cosineSimilarity(query, result.embedding);
+ const keywordScore = exactMatchBoost(query, result.title);
return semanticScore + freshnessBoost + keywordScore;

Book a private demo

Mezure is onboarding paid pilots with engineering teams hiring in the agentic era. Share your work email and we'll follow up to scope a pilot around your hiring process.