An open specification (Apache-2.0) for making "verified" a testable claim about a model. Donated so it can become the default; certified so it means something.
The count of observables that are simultaneously measured, computable and validated — with the per-observable verdict published, not just the total.
r(δ) = sqrt(ln(1/δ)/(2m)). When m = 0 the honest answer is "extrapolation, not verification".
p* ≤ p̂g + p̂e + rs + re — a model is never "verified" merely by matching another model.
Named checks; a skipped check is a fail; every proven guard has a recorded run where it went red.
Who produced it, that it ran, what falsifies it. A result missing any of the three is unverified.
L1 Gated · L2 Proven · L3 Certified — the mark is L3 only.
The conformance suite is public and it is designed to be able to fail. A fresh clone runs in seconds:
git clone https://github.com/robmitchellengineering-hub/vds-spec
cd vds-spec/conformance
python3 vds_conformance.py --candidate candidates/reference.json # PASS
python3 vds_conformance.py --candidate candidates/nonconforming.json # FAIL (by design)Reference: pacland-climate — m = 4 of 12, radius 0.6791, all six checks PASS. The broken candidate is rejected with the exact failures named.
An adopted standard is the one anyone can implement and anyone can test. The specification is Apache-2.0; the certification is the audited service. That is the SOC 2 / UL shape — donate the rule, audit the conformance.
A Technical Steering Committee owns the text, the suite and the certification rules, with VVS at most 2 of 7 seats. Open RFCs, public decisions. Until the committee is seated with independent members the standard is described as an open draft by VVS — and it says so.