DX-Score and certification

DX-Score General suite

Results from different test-suite versions must not be mixed in one leaderboard comparison. The tested method version, test-suite version, and leaderboard baseline version remain attached to the record so a later baseline does not silently rewrite a historical result.

Case-level 1–5 rubric

Every case receives a score from 1 to 5. Higher is better.

Calculate the DX-Score

The module means and combined score are calculated as follows:

bench_dx_mean =
  sum(score for each of 3,000 BenchDX cases) / 3,000

robust_dx_mean =
  sum(score for each of 3,000 RobustDX cases) / 3,000

dx_score = bench_dx_mean + robust_dx_mean

• BenchDX is displayed with two decimal places as x.xx / 5.
• RobustDX is displayed with two decimal places as x.xx / 5.
• Their sum is displayed with two decimal places as x.xx / 10.

Read the score and leaderboard

The public leaderboard ranks concrete model versions by the combined DX-Score. It does not rank a model family or organization as though every release behaved identically.

DX-Score

The sum of the BenchDX and RobustDX module means. It is the default leaderboard ranking value.

BenchDX mean

The average 1–5 rubric score across the fixed 3,000 BenchDX cases.

RobustDX mean

The average 1–5 rubric score across the fixed 3,000 RobustDX cases.

Tested version

The concrete model or system release represented by the row. A later version requires its own result.

Locked baseline

The named model cohort and associated versions used to interpret rank and certificate qualification. Current DX-Score product copy describes the cohort as including open-source and commercial models; the recorded baseline version determines its exact membership.

Certification levels

Qualification is relative to the locked comparison cohort recorded by the public-model leaderboard baseline under the same DX-Score General method version. Current DX-Score product copy describes this cohort as including open-source and commercial models.

“P80” and “top 20%” describe the same boundary from opposite directions: a result at or above the 80th percentile has outperformed at least 80% of the comparison cohort and falls within its top 20%.

From test to public record

1. 01

   Define the tested scope

   Lock the subject identity, version, configuration, access method, and intended deployment boundary before testing.

2. 02

   Run the fixed suite

   Execute all 6,000 cases under the declared DX-Score General method and test-suite versions.

3. 03

   Review the evidence

   Complete scoring and risk-based review, and preserve the detailed findings in the private test report.

4. 04

   Compare the locked cohort

   Compare the result with the applicable locked model cohort under the recorded leaderboard baseline.

5. 05

   Publish the verification record

   A qualifying result receives a badge level and public verification identifier tied to its tested scope and version locks.

What the verification record shows

The public verification surface is intentionally narrower than the private evaluation report. The current public implementation shows the following certificate facts:

DX-Score General records use the package DX-Score General, the modules BenchDX + RobustDX, and the issuer DX-Score by AIDX.

Certificate status is separate from badge level. A public record can indicate that a certificate is active, suspended, revoked, or superseded. Verify the current public record instead of relying on a copied badge image.

Certificate scope and limitations

A DX-Score badge is a certificate of testing. It verifies that AIDX tested the named subject and version under DX-Score General and issued the recorded badge level under a locked method, test suite, baseline, and declared scope.

Next steps