The Atlas shows where a research portfolio's public evidence chain breaks—and which identifiers to fix. In a full audit, only 17 of 645 EU–Japan quantum publications expose project/funder, dataset and software links in OpenAIRE.
Working artifact · 645/645 records audited · inspect sources · export JSON, CSV or Markdown
01 · THE EVIDENCE GAP
“Connected” means at least one explicit OpenAIRE edge. It is a Graph observability signal—not proof of access, licensing, documentation quality or actual reuse. “Not connected” means only that the edge was not observable. A “complete chain” is publication-centred; it does not assert that a named grant caused or produced the connected dataset/software records.
02 · ONE COMPLETE CHAIN
Graph-linked funding
The publication carries three result→project edges. Each project edge contains an EC funder object—more reliable here than the top-level publiclyFunded flag.
OpenAIRE recordFeatured chain in the 29 Jul 2026 census: OpenAIRE product doi_dedup___::82d884… · Later drift is disclosed below
03 · TWO-LAYER OPENAIRE DESIGN
The Alien/OpenAIRE MCP is the interactive verification layer; the direct Graph API is the deterministic census layer. The 30-call MCP trace followed the featured DOI and grant, while a targeted API follow-up isolated a zero-based pagination mismatch without allowing that connector defect into the 645-record denominator.
“Using the OpenAIRE Graph only, cross-check DOI 10.1088/1367-2630/ad5b13. Return identity, funding, projects, organisations, datasets, software and provenance. Treat missing relations as not observable.”
DIRECTLY OBSERVED · 18 JUL MCP RUN
doi_dedup___::82d884…ROOT CAUSE CONFIRMED · OFF-BY-ONE
The official link endpoint defaults to page 0. Direct calls returned all three typed rows there; pages 1 and 2 returned none. Alien's ResearchLinksInput rejected page 0 before an upstream call. This is an MCP schema/offset defect, not missing OpenAIRE data.
Time-bounded result: the table records the 18/29 July diagnostic. On 20 August the featured dataset total remained 0, while the two software rows still reproduced the page-0 mismatch. MCP remains the inspectable traversal layer; the API remains the deterministic census layer.
04 · OBSERVABILITY TIMELINE
Q‑NEKO is a new EU–Japan joint quantum-technology project: €4m in European funding, running 2026–2028. It was absent from the 29 July queries. By 20 August, the bounded recheck found one exact project record and five raw product hits. Two were this Atlas's own Zenodo releases; after self-exclusion, the remaining three records also resolved through the explicit Q‑NEKO project-code relation.
The clock produced its first state change.
The Atlas detected project visibility ten days after its published census; aggregate rates remain the 29 July snapshot pending an authenticated refresh.
05 · DECISION BRIEF
One finding, three operating moves.
Use a declared denominator to see where funding, dataset and software visibility falls—before interpreting project performance.
Require grant IDs, publication DOIs, repository IDs and software identifiers before project close—not after evaluation.
Schedule the same bounded audit again. Q‑NEKO demonstrates how to record Graph-entry milestones without turning a missing edge into a performance verdict.
If added repository identifiers or improved OpenAIRE classification lifts software visibility near the 28% dataset rate, the “software cliff” weakens. This Atlas publishes its denominator so that claim can be tested.
The 29 July full census and 20 August bounded probe answer different questions. The first fixes the aggregate denominator; the second detects Graph drift without silently rewriting the published rates.
A publication-count view asks how much research is visible. The Atlas asks whether a decision-maker can audit the whole path from investment to reusable outputs, preserves missing-edge uncertainty, and turns a connector failure into a regression test.
06 · REUSE THE AUDIT
The artifact separates observed links from unknowns, retains API URLs in the analysis cache, and publishes source attribution, transformations and licence metadata. Snapshot, notebook and the 23-request live probe are credential-free; a fresh full census requires an OpenAIRE bearer token and is throttled below the authenticated limit.
Discover with eight quantum phrases; deduplicate the Japan-query union.
Resolve scope to 2020–2026 publications with Japan + EU27 affiliations.
Audit all 645 for project and funder objects in product links.
Audit Scholix links for all 645 observed records; retain the 87 title-literal sensitivity set.
Compare subsets with Wilson 95% binomial reference bands.
REPRODUCIBILITY PACK
Evidence snapshotJSON ↓Connection ratesCSV ↓Executed analysisIPYNB ↓Analysis reportMD ↓Source pipelinePY ↓Executed MCP cross-checkMD ↓MCP run manifestJSON ↓MCP pagination reproducerPY ↓Bounded live recheckJSON ↓Full-census verificationJSON ↓Live recheck sourcePY ↓Three-minute verification guideMD ↓Responsible-use manifestJSON ↓FAIR RO-Crate metadataJSON-LD ↓1–2 page storyMD ↓Captioned walkthrough119 SEC ↗OpenAIRE endpointAPI ↗Full census snapshot: 29 Jul 2026 · latest bounded recheck: 20 Aug 2026 · fresh full census requires authenticated access