OPEN SPECIFICATION

PAVS — Open Specification for Physical AI Verification & Scoring

PAVS defines how to score physical AI in simulation, how to verify that scoring, and how to report the result.

The standards stack — the middle is empty

Standards exist above and below. What is missing is the middle layer: where a value came from in simulation, and whether a verdict was established at all.

The standards above and below presuppose physical testing; measurement, verdict and reporting in simulation fall outside that premise.

Five requirements for trustworthy scoring

PAVS defines scoring as a chain of five stages and defines the requirements of each axis as 72 test assertions. A test method is specified for all 72 (injection 36 · inspection 33 · demonstration 3). Conformance is determined from the records of tests performed under this procedure.

1. 1

   Test items

   What is measured

2. 2

   Measurement

   Where the value came from

3. 3

   Verdict

   The pass criterion

4. 4

   Verification

   Evidence behind the verdict

5. 5

   Reporting

   How the result is stated

Unmeasured is not a pass

Scoring has three outcomes, not two. If the measurement or stimulus an item requires never took place, the result is neither a pass nor a failure.

Not measured is never counted as a pass. A missing measurement for a mandatory item is treated as a fail with the reason “measurement absent”, and the number of unmeasured items is kept separately. A value that appears as 0 or “true” has not thereby been measured. (PAVS-VDT-01 v1.2)

The authors tested themselves first

On 4 September 2026 we tested the imSim scoring system against v1.2 of this specification. The results are published as they are. The current version, v1.4 (2026-10-19), is an editorial revision with the same assertion requirements.

The specification forbids claiming conformance while any core assertion is unmet. The imSim scoring system is therefore described as "PAVS-aligned", not "PAVS-conformant". All 65 unmet assertions are listed individually in the conformance statement. The test was carried out by the developing organisation itself and is not a third-party audit; the conditions that would invalidate it are stated in the statement.

The artefacts

What follows is not illustration. These are real outputs of the system.

Domain profiles

The specification body is domain-neutral. Which items each domain measures and which it excludes are published as a profile. The first profile covers the military domain.

Documents

The specification documents are currently published in Korean only.

• White Paper (Part 0) v1.2informativeThe background and design rationale of this specification. Covers the precedent of simulation-based validation institutionalised for autonomous driving, the gap in robotics, and the publisher's self-test results.PDF
• Specification (Part 1) v1.4normativeRequirements and conformance. Specifies 72 test assertions and the test method for each. v1.4 = an editorial revision with the same requirements as v1.2.PDF
• Casebook (Part 2) v1.2informative39 incidents from which clauses arose. Actual records of scoring that went silently wrong, with no error shown, are given as the basis for the clauses.PDF
• Conformance Statement v1.2self-declarationResults of testing the imSim scoring system against this specification (test date 2026-09-04). Every met and unmet item is disclosed.PDF
• Reference Implementation (Part 3) v0.5informativeThree demonstrations that execute the specification's test assertions in an actual implementation, and the composition of the run package. An assertion shown to be executable is not thereby judged as met.PDF
• Military profile v0.6Domain profileWhich items are measured and excluded in military scenarios, the verdict mapping, the self-check and limits, with the results of the first tests (8 maritime encounters) as an annex. Based on Part 1 v1.3.PDF