© Ribose Asia Limited 2026
This document is not a Ribose Standard. It is distributed for review and comment, and is subject to change without notice and may not be referred to as a Standard. Recipients of this draft are invited to submit, with their comments, notification of any relevant patent rights of which they are aware and to provide supporting documentation.
All rights reserved. Unless otherwise specified, no part of this publication may be reproduced or utilized otherwise in any form or by any means, electronic or mechanical, including photocopying, or posting on the internet or an intranet, without prior written permission. Permission can be requested from the address below.
Ribose Asia Limited
Suite 1, 8/F, New Henry House
10 Ice House Street
Central
Hong Kong
www.ribose.com
This part is the gate of the series: it defines the corpora that freeze normative behavior into replayable vectors, the format those vectors take, the harness every implementation must satisfy, and the conformance classes an implementation may claim. Byte-identity across implementations is enforced HERE, by replay — never by shared code.
Specifications without gates are aspirations. Every clause of Parts 2 to 7 is paired with vectors that a machine can replay, and this part defines the machine: the layout, the format, the tolerance convention, the classes. An implementation’s claim is exactly the set of vectors it replays — no more, no less.
This document specifies: the corpora layout and versioned index; the vector file format and metadata; the replay-harness requirements; the runtime conformance classes and the claims each licenses; and the reproducibility classes for Monte Carlo vectors.
No terms and definitions are listed in this document.
one test case: an input, an expected result (exact decimal strings per CalConnect 36813, and metadata.
any byte difference between a replay and its vector, beyond a vector’s declared tolerance.
the execution of a vector by an implementation under test, compared against the expected result.
a named set of vector classes (Clause 8) licensing specific claims.
byte-exact or distributional, per Clause 9.
Requirement language per CalConnect 36812. Tolerances are written tolerance_d with |computed − expected| ≤ 10−d·|expected|; vectors without tolerance_d are byte-exact.
The index is an artifact (CalConnect 36812, artifact discipline): checksummed, editioned, consumed by both reference and parity runtimes.
A vector carries: input; expected as exact decimal strings (CalConnect 36813 for every compared field; metadata: the normative source (document + clause), engine_version, grammar_version, vector_set_version, numeric_algorithm, and — only where a vector is non-exact — tolerance_d. Monte Carlo vectors carry the PRNG identity, seed and trial count.
Expected shape. The degC vector: input (20 °C).convert_to(K); expected "293.15"; metadata citing CalConnect 36816 Clause 5, the engine and vector-set versions, and no tolerance (byte-exact).
A conforming harness: load and checksum-verify the corpora; replay every vector of every claimed class; compare bytes; and report drift per vector. Release CI blocks on any drift. The reference harness lives in unitsml/unitsml-testsuite; parity runtimes (TypeScript, Rust) run the same corpora through their own lanes.
Implementations are licensed to claim conformance as in
| Class | Licenses the implementation to claim |
|---|---|
| parse/render | correct parse of an input language into the model and correct emission per format column of the render corpus |
| quantity-conversion | CalConnect 36815/5 semantics: resolution, gating, scales, conversion application (Annex B) |
| evaluation | Plurimath 4: propagation, adequacy verdict, second-order, MCM, conformity (Annex C) |
| full-series | all of the above plus round-trip guarantees (Plurimath 2, Annex A) |
Claims name the class and the corpus version replayed; an implementation never claims a class whose vectors it does not replay.
Expected claim shape. “Conforming at class quantity-conversion, corpus v2026.10” licenses resolution, gating, scales and conversions — and says nothing about evaluation.
Byte-exact: vectors produced by the pinned Wichmann–Hill generator with recorded seed and trial count; the adaptive stopping δ = ½·10^ℓ is part of the vector. Distributional: for methods without deterministic draws (e.g. MCMC sampling, GUM-5 Ex 12), vectors carry statistical-equivalence assertions (stated test and threshold) instead of byte expectations. Implementations state which class each claimed vector satisfies.
Expected outcome. The JCGM 101 §9.5 vector records Wichmann–Hill with its seed and trial count; a runtime producing 36 nm-class results with a different generator claims the distributional class, not byte-exact.
This part is the gate of the series: releases across the family are blocked on vector drift, and conformance claims are exactly the replayed classes of Clause 8. The frozen vector sets are given in
Parse/render/round-trip vectors for every mechanism of Plurimath 2: the three categories; k, p pairings; interval kinds; concise (output-only) and absolute notations; the ASCII channels (quoted channel for dialect L; native literals for pma); lossiness assertions per the Plurimath 2 matrix. Invariant replayed: every model field survives every lossless channel byte-identically.
Round-trip triplet. ucov(10.0, 9.7, 10.3, 0.95, shortest) through MathML → model → OMML → model preserves category, endpoints, p and kind byte-identically; the LaTeX column renders the visible interval only, per the declared loss.
The vector tables of CalConnect 36815 Annex A and CalConnect 36816 Annex A, frozen: degC/degF round-trips, percent, non-exact percent-per-degree, difference-skip-addends, affine-in-composite and power ≠ 1 errors, ft² double-power red/green, refusals, resolution red/greens (J → energy; Bq↔Hz; equivalence-class addition), scale errors naming the class.
GUM H.1–H.6, with H.1 as the versioned triple: 2008 first-order (u_c = 32 nm) / amended second-order per Amd.1:2026 H.1.7 (u_c ≈ 34.2 nm; derivation in Plurimath 4 Annex A) / MCM per JCGM 101 §9.5 (u(δL) = 36 nm; shortest 99 % interval [745, 932] nm).
JCGM 101 §9.2 (masses), §9.3 (microresistance), §9.4 (comparison loss with r = 0.9), §9.5 (gauge block) MCM vectors; pinned Wichmann–Hill seeds.
LPU-failure set: GUM-6 §13.1.4 bromine (LPU u = 0 vs true 0.00007); GUM-5 Ex 6 (dominant uniform), Ex 7 (GUF₁/GUF₂/MCM), Ex 13, Ex 14 (linearity adequacy), Ex 16 (comparison loss at x₁ = 0).
JCGM 106 decisions: one-sided p_c at Table 1 z-values; guarded accept/reject with r = 0.65 → R_C = 0.1 %; C_m = 1 boundary; ISO 14253-1 w = U ceiling.
Every vector names its normative source. The H.1 triple’s second-order member carries the derivation reference (Plurimath 4 Annex A); the JCGM vectors cite clause and table. The excluded-as-byte-vectors set (MCMC draws) is listed with its distributional-class alternative.
The following documents are referred to in the text in such a way that some or all of their content constitutes requirements of this document. For dated references, only the edition cited applies. For undated references, the latest edition of the referenced document (including any amendments) applies.