phionyx/measurement-axioms-cases

Measurement Axioms Cases

Source pinning

Frozen, source-pinned publication — not a live mirror of the canonical repository's main.

Source snapshot commit	350bb4cba4e5bc2d760db080aae52352a7041331 (Measurement Axioms v1.0.0)
Canonical current repository	https://github.com/halvrenofviryel/measurement-axioms
Export/publication date	2026-09-13 (first Hub commit of this repository)
Update policy	Counts are derived from this snapshot: 45 active requirements (15 of them assessable: review, kept NOT_MEASURED by automation), 14 generated cases, 14 reference runs; the 314 upstream unit tests are software tests, not cases.

Later commits on the canonical repository are not reflected here unless a new publication says so. Byte identity with the snapshot is a distribution-integrity property, not independent verification.

A public reference/conformance corpus, not a blind benchmark or an estimate of population-level system performance. This is a distribution of the frozen Measurement Axioms v1.0 source, commit 350bb4cba4e5bc2d760db080aae52352a7041331. GitHub remains canonical. The Hugging Face mirror is not a new independent study.

Units and actual inventory

Config	Rows	Unit
requirements	45	Active normative requirement parsed from frozen YAML
cases	14	Generated local fixture input, one side of a pair
reference_runs	14	Actual local execution of a case

There are 7 intact control/degenerate pairs across six scenario families. The train name is the Dataset Viewer split convention; no learning/evaluation split or random pair-breaking split is implied. The upstream 314 software test instances are not 314 benchmark cases. Fifteen requirements require human review and remain NOT_MEASURED in automated assessments.

All current cases have case_origin: generated. Fixed local demonstrations derive from the cited normative vectors and source test symbols; they are not historical field incidents. input_payload is serialized JSON. fixture_hash hashes those exact UTF-8 bytes. case_id includes pair, role, source, input, expected status and scope. requirement_ids and vector links indicate relevance, not a claim that every associated requirement was fully assessed.

Expected versus observed

cases.expected_* contains declared expectations. reference_runs.observed_* is populated only by running the frozen reference constructors/schema and fixed local targets. test_passed compares expected and observed results; a successful software test can observe NOT_MEASURED or FAIL. The reference's pair probe only assesses its stated generated target/pair. It is not an overall conformance verdict. Immutable outputs live in runs/, their environment in environment.json. implementation_commit identifies the reference; adapter_source_hash separately identifies the non-normative adapters, and adapter_source_commit records the actual source revision (adapter_worktree_modified discloses uncommitted code). output_hash hashes canonical publication JSON for the actual output, not an external attestation.

Scope and limitations

Record schema validation establishes shape only. The local evidence registry, attenuation and constructor replay are synthetic demonstrations. No arbitrary URL is fetched; no external reference, intervention, identity, signature, key custody or independent observer is verified. Static analysis is not supplied. Review controls do not become PASS automatically. A populated chain fixture only checks its records' schema, not cryptographic links. These narrow checks do not establish organization-level or system-wide conformance, model quality, safety, ranking or certification. No LLM judge or inference is used.

Reproduction and pinned loading

The source adapter is publication/measurement/build.py in the linked Space's development repository; distribution includes unchanged frozen reference files under source/, source/license hashes, actual raw outputs and exact runtime files under adapter/. source-lock.json records source identity, MANIFEST.sha256 checks distribution bytes. Neither hash proves an external measurement occurred.

To repeat these fourteen actual runs locally from a dataset snapshot: python reproduce.py (requires Python 3.12, jsonschema 4.25.1 and PyYAML 6.0.3). The script runs the included adapter/reference and compares actual output hashes; it does not copy expected labels into observed results.

from datasets import load_dataset
data = load_dataset("phionyx/measurement-axioms-cases", "cases", revision="<full HF commit>")

Use a real full revision after publication. Each config explicitly selects one JSONL file; source files and raw runs are not additional Dataset Viewer rows.

Attribution and license

Ali Toygar Abak, The Measurement Axioms, v1.0.0, CC-BY-4.0. Concept DOI, frozen v1.0.0 DOI. The source LICENSE and CITATION.cff are preserved byte-for-byte. This adaptation adds a dataset index, generated fixtures and scoped UI; frozen specification, requirements, schema and reference implementation are unchanged. The license does not grant a right to claim conformance.