Spherity Research · Publication overview and final PDFs

Proofs over Statements: Verifiable Evidence Graphs

Credential formats for European Business Wallets, industrial AI and mixed-assurance evidence

Author
Affiliation
Dr. Carsten Stöcker — Spherity GmbH
Published
Updated
Research cut-off
· Final editorial review; individual references retain the verification dates recorded in the paper
This version
https://spherity.github.io/spherity-research/proofs-over-statements-evidence-graphs.html
Latest version
https://spherity.github.io/spherity-research/proofs-over-statements-evidence-graphs.html
Browse this paper

In brief

What does this research establish?

For industrial evidence that is combined across organisations, retained and repeatedly verified as graph data, this paper recommends a common W3C VCDM 2.0 JSON-LD profile secured by RDF-based Data Integrity. Preserve source snapshots, interpretation rules and assurance levels; add an envelope when a defined presentation or interoperability requirement justifies it. This is a conditional engineering recommendation, not a claim that every nested credential is insecure or that graph proofs are always faster.

Key takeaways

  • JSON-LD syntax alone is insufficient: verification needs the authenticated proof inputs, controlled interpretation and recoverable source boundaries.
  • Envelope-only B1 and a nested credential retaining an inner Data Integrity proof have different verification obligations; the paper distinguishes them.
  • Ordinary EAAs and policy-required QEAAs should be able to use a common graph profile while retaining their different assurance levels and issuance requirements.
  • The prototype cost advantage depends on repeated verification with refreshed envelope-to-graph mapping. Protected cached mappings and presentation workloads can favour other paths.
  • The release supplies complete signed plain and nested examples, exact tokens, public verification material and passing positive and negative test reports.

Final publication downloads

This is the research overview for the final publication package, version 1.3, 9 November 2026. The PDFs are the authoritative full paper and executive brief.

The ZIP contains both PDFs, editable LaTeX sources, vector figures, complete signed credential examples, exact tokens, public verification material, pinned dependencies, validation reports and checksums. Each document folder is self-contained. The executive brief introduces business-wallet and industrial requirements for readers who already understand credentials; its appendix gives the same complete verification examples as the full paper.

What must remain verifiable after evidence is merged?

A citizen may present selected attributes to a verifier. An industrial workflow can retain and combine evidence from manufacturers, suppliers, laboratories, authorities and automated systems. Battery-passport analytics, customs review and an aerospace safety supervisor are the paper’s worked examples. These workflows need the meaning, issuer, source version and applicable authority of an assertion to remain inspectable after it becomes part of a larger graph.

The architectural question is therefore not simply whether a credential contains JSON-LD. Both a graph-proof credential and a nested envelope can carry JSON-LD. The question is which inputs the issuer authenticated, what the relying party preserves, and how a machine binds the fact it uses to that authenticated evidence. Shared identifiers and RDF semantics help combine facts; credential proofs, source boundaries and decision policy address different parts of their use. The full paper develops these distinctions in Sections 4–5.

A citizen presents selected credentials while organisations combine EAA-backed evidence with selected policy-required QEAA relationships in a common graph stack.
Figure 1. Citizen presentation and retained business evidence. Green relationships are EAA-backed; blue dashed relationships require QEAAs under the decision policy. The proposed common graph stack preserves issuer, scope, status and assurance level. Colour denotes supporting attestations, not independently signed edges. The examples illustrate workloads, not exclusive wallet capabilities. Source: executive brief, Figure 1, final publication package v1.3.
Download figure: SVG · PNG · PDF

A common graph profile, with envelopes where they serve a purpose

The paper recommends W3C Verifiable Credentials Data Model 2.0 expressed in JSON-LD, secured with an RDF-based Data Integrity suite, as the persistent evidence format for the specified retained industrial graph workload. A profile is more than a serialization: it also fixes interpretation rules, admissible proof suites, source recovery, issuer/key binding, status handling and application validation. W3C VCDM 2.0 supplies the data model; Data Integrity supplies proof mechanisms.

Construction What the verifier must establish Consequence for retained graph evidence
A: RDF-based Data Integrity The suite’s authenticated credential dataset and proof configuration, with authorised verification material and profile checks. Recover the signed source snapshot, original graph positions and proof inputs. An arbitrary merged graph with source boundaries erased is insufficient.
B1: JSON-LD in an envelope, without an inner graph proof The protected token input and disclosure commitments, plus the governed interpretation and binding to the graph view. Keep the relevant source inputs available or reliably reconstructible. A protected cached mapping is a valid design when its binding, freshness and status are governed.
Dual-proof nesting in the complete examples The outer ES256 signature and disclosure commitment, and the inner Ed25519 Data Integrity proof; both must pass under the example profile. The inner graph-proof route is preserved, with additional obligations whenever policy relies on both layers. These are not the B1 benchmark timings.

Use an envelope for a defined presentation, disclosure or interoperability requirement. A presentation-only wrapper can be checked at ingestion, followed by direct verification of the retained inner credential when policy permits. Its extra proof work need not recur on every graph read. Equally, selecting a graph format does not eliminate conflicting application schemas, unsafe processing, stale evidence or false statements by an authentic issuer.

What the cost and assurance results establish

For the paper’s 146-quad synthetic fixture, the additional stored graph-proof check took about 0.28 ms, compared with 1.06 ms for envelope processing with fresh JSON-LD expansion. Including issuance and ingestion, the model favoured the graph path from the fifth additional verification. The separately measured envelope-only check took 0.25 ms, excluding validation of the cached graph view and its source binding. These are prototype processing paths, not full service latencies or universal format rankings. A separate presentation comparison favoured the tested SD-JWT implementation under its different proof procedure.

The policy may accept ordinary electronic attestations of attributes (EAAs) for product data, events and evaluations while requiring qualified electronic attestations of attributes (QEAAs) for selected foundational relationships. Sharing a credential profile does not erase their assurance differences. Qualification requires an eligible issuance profile and the applicable provider, scope and lifecycle conditions; it does not automatically qualify a neighbouring claim or AI output.

If a required QEAA is missing, the affected justification path cannot meet the target assurance. Another admissible path may still satisfy policy. If no such alternative exists, the decision lacks the required assurance, while independent evidence remains verifiable. The attack model additionally considers genuine-but-false content, changes to authority and joint attacks. The standards recommendation is to make an appropriate graph-proof profile eligible alongside existing options, rather than treating a wrapper as sufficient qualification. The paper discusses the European Business Wallet proposal as a proposal and distinguishes it from the adopted eIDAS amendment.

Complete signed examples and practical verification steps

Full-paper Appendix C and executive-brief Appendix A provide an actual signed VCDM credential and the same credential disclosed inside an SD-JWT. Both assert an illustrative battery’s recycled-cobalt share of 16 percentage points. Public keys, contexts, proof values, the complete wire token and validation scripts are included in the ZIP; no private signing key is needed.

  1. Verify the plain credential. Enforce its profile, resolve the pinned context and authorised issuer key, prepare the credential and proof-configuration inputs, canonicalise them, hash them and verify the Ed25519 signature. Expose only the verified view to the application.
  2. Unnest and verify the outer layer. For the nested example, first verify the ES256 issuer signature and allowed header/claims, match the disclosure hash, reconstruct the credential and enforce the inner/outer issuer binding. Decoding alone authenticates nothing.
  3. Verify the recovered inner credential. Apply all plain-credential checks. A saved negative fixture has a valid outer signature and disclosure commitment but an invalid inner proof; the combined verifier rejects it.
Work or representation Plain credential Nested credential
Signature verification One Ed25519 One ES256 plus one Ed25519
RDF canonicalisation Two suite inputs The same two inner inputs
Explicit SHA-256 work Two inner hashes Two inner hashes plus one disclosure hash
Disclosure reconstruction None One whole-credential disclosure
Serialized size 797 bytes of compact UTF-8 JSON 1,534 bytes of combined SD-JWT

The size comparison excludes external key and context documents; the explicit hash count excludes hashing internal to signatures and canonicalisation. The whole credential is disclosed as one unit, so the inner eddsa-rdfc-2022 proof does not supply selective disclosure of individual battery attributes.

The main test runner passed four acceptance controls and 13 expected rejections. A second audit passed nine assertions and eight negative checks through direct Node/OpenSSL signature verification. It shares the RDF processing dependencies and is not an independent RDF implementation. These local tests establish the examples’ verification behaviour, not legal qualification, live revocation, wallet interoperability or the truth of the assertion.

Connections to Trusted AI, industrial AI, physical AI and PQC

Knowledge graphs express what entities and relationships mean. Verifiable evidence graphs preserve authenticated assertions, sources and versions. Assurance graphs connect system claims to evidence, requirements, assumptions and reasoning. The paper argues that these complementary views belong in the core evidence infrastructure for efficient, safe and compliant AI systems, alongside the engineering and governance needed to achieve those outcomes.

For Trusted AI, this makes the supporting evidence and authority behind a proposed action inspectable. For industrial AI, it links products, process conditions, tests and compliance requirements across organisations. For physical AI, such as robots and autonomous machines, it can connect permitted actions and operating limits to current evidence and authorised actors. Cryptographic attribution supports those decisions; it does not replace content validation, monitoring, runtime safeguards or a designed safe response when verification is unavailable.

Cryptographic hybrids and nested credential formats are different constructions. Data Integrity already supports proof sets and proof chains. Combining traditional and post-quantum signatures still requires validated suites, key and credential bindings, verification of mandatory components and downgrade protection. A common profile can reduce integration variation, but additional signatures do not elevate an EAA to a QEAA. The paper’s performance measurements do not include PQC overhead; its treatment follows the distinctions in Hybrid Signature Spectrums, RFC 9955.

Figures in SVG, PNG and PDF

The gallery presents the figures from both final documents. Each caption identifies its source and the scope of the illustrated result. Download SVG for scalable vector artwork, PNG for presentations and web use, or PDF for publication-quality placement. Web SVG lettering is outlined for reliable display; the publication ZIP retains the original editable SVG sources. The introductory citizen-and-graph comparison above is Figure 1; the remaining figures follow below. Credit Carsten Stöcker, Spherity GmbH, link to this research page and CC BY 4.0, and indicate changes.

Results, assumptions and worked designs connect the RDF encoding, proof recovery, reuse cost and assurance arguments.
Figure 2. Structure of the argument. Rounded boxes are results. Grey boxes identify modelling conditions and premises. Arrows show how the results and worked designs are related. Source: full paper, Figure 1, final publication package v1.3.
Download figure: SVG · PNG · PDF
Three layers connect schema.org JSON-LD, verifiable evidence graphs and the proposed graph-proof profile for qualified authority attestations.
Figure 3. Web markup (schema.org in JSON-LD), a verifiable evidence graph and qualified authority edges share one data model. The qualified layer shows the graph-proof profile proposed in Section 9. Source: full paper, Figure 2, final publication package v1.3.
Download figure: SVG · PNG · PDF
Architecture A authenticates canonical RDF quads and proof configuration; architecture B authenticates envelope bytes and needs recoverable source evidence after a merge.
Figure 4. The architectures can convey the same RDF dataset. A checks a proof over the canonical quads and the canonical proof configuration. B checks a proof over the protected envelope input. In the erasure case of L2(b), B needs retained source evidence for direct verification after a merge. Source: full paper, Figure 3, final publication package v1.3.
Download figure: SVG · PNG · PDF
Cumulative and per-check prototype costs compare graph proofs, fresh envelope-to-graph expansion and an envelope-only check with a trusted cached mapping.
Figure 5. Cost model using run-3 prototype timings for IRI-only credentials on one thread. (a) The curves sum median batch means for issuance, ingestion and R re-verifications of a 146-quad credential. The calculated crossover is R*=4.24 for the envelope path with fresh expansion: at that point both paths have accumulated about 6.1 ms; left of it B costs less because its issuance and ingestion are cheaper, right of it A costs less because each further check costs 0.28 ms in A and 1.06 ms in B, so that at R=10 B has cost 1.6 times A. The dashed envelope-only curve excludes checks on a cached RDF view. (b) Measured per-check costs A₁, B₁ and B₁ˢ by credential size. The crossover labels use B1 and include issuance. Source: full paper, Figure 4, final publication package v1.3.
Download figure: SVG · PNG · PDF
Qualified authority checks do not eliminate a feasible independent attack on an ordinary content declaration that changes the accepted outcome.
Figure 6. An illustrative policy under A2. (a) An authority edge validated at E bounds the decision by K(E) if it is a single direct attack target satisfying T3(c). (b) If graph-profile issuance and validation make those edges Q, an E declaration that can independently change the accepted outcome still bounds the decision. Equality requires a matching lower bound from T3(a). Source: full paper, Figure 5, final publication package v1.3.
Download figure: SVG · PNG · PDF
A battery passport joins issuer declarations, verification statements, digest links and authority checks for a carbon-footprint reporting decision.
Figure 7. Evidence graph behind one battery passport. Solid grey arrows are digest links. Dashed arrows show authority checks. The orange path shows the declaration and verification statement used by the reporting decision. The example applies the decision rule in Section 6.1. Source: full paper, Figure 6, final publication package v1.3.
Download figure: SVG · PNG · PDF
Authority facts, signed shipment evidence, digest links and release-decision inputs form a cross-organisation customs evidence graph.
Figure 8. Evidence graph of one e-commerce shipment at the port of entry. Blue boxes are authority facts, green boxes issuer-signed evidence, solid arrows digest links, dashed arrows authority checks, orange arrows the evidence used by the release decision. The data model is reconstructed from the public pilot demonstration. Source: full paper, Figure 7, final publication package v1.3.
Download figure: SVG · PNG · PDF
A safety supervisor checks proposed actions against authenticated requirements, records its decision and keeps local protective functions separate.
Figure 9. Safety supervisor consuming a verifiable requirements graph. The supervisor evaluates proposed actions against authenticated requirements and records the outcome. Supporting credentials are checked at load and update, and again per decision if the policy requires it. Local protective functions operate separately. Timing examples come from the prototype measurements and the projections in Tables 6 and 7. Source: full paper, Figure 8, final publication package v1.3.
Download figure: SVG · PNG · PDF
Side-by-side steps compare stored graph-proof verification with envelope-only B1 processing, disclosure reconstruction, expansion and source binding.
Figure 10. A machine checking the same stored battery claim through A and envelope-only B1. The dual-proof nested example in Appendix A additionally verifies the inner graph proof. Unnesting includes authentication, disclosure reconstruction and controlled interpretation; decoding alone proves nothing. The displayed timings cover prototype portions of the process, not all steps or complete security checks. A cached mapping can replace fresh expansion only under a policy protecting its source binding, freshness and status. Based on the full paper, Sections 5.2–5.5; explanatory diagram added in this edition. Source: executive brief, Figure 2, final publication package v1.3.
Download figure: SVG · PNG · PDF
Measured prototype times compare a stored graph-proof check with envelope processing and fresh RDF expansion; the caption distinguishes cached-view checks.
Figure 11. Prototype times per additional verification, by credential size. The orange path verifies the retained envelope and freshly expands its payload; comparison with a cached graph is outside that timed path. The blue path retrieves and verifies the stored credential and proof inputs. At 146 quads, the separately measured envelope-only path took 0.25 ms, excluding validation of the cached view and source binding. Based on the full paper, Table 3. Source: executive brief, Figure 3, final publication package v1.3.
Download figure: SVG · PNG · PDF
The graph-proof path costs more initially and crosses the fresh-expansion path at 4.24 additional checks for the 146-quad fixture; the cached-envelope path has a different scope.
Figure 12. Cumulative processing and per-check costs from the full paper, including the envelope-only alternative. The graph path costs more initially and crosses over against fresh expansion at 4.24 additional checks for the 146-quad fixture. The dashed envelope-only curve assumes a trusted cached mapping and excludes its validation. These are prototype processing costs, not complete service latencies. Based on the full paper, Figure 4. Source: executive brief, Figure 4, final publication package v1.3.
Download figure: SVG · PNG · PDF
The battery-passport graph illustrates why a valid mandate or qualified authority does not itself authenticate the exact reported carbon-footprint class.
Figure 13. The battery-passport design distinguishes content, evidence links and authority checks. A valid mandate and accredited issuer do not by themselves authenticate the exact reported carbon-footprint class. The diagram includes a content attack that can remain possible even when authority attestations are qualified. Based on the full paper, Figure 6. Source: executive brief, Figure 5, final publication package v1.3.
Download figure: SVG · PNG · PDF

Evidence, reproducibility and limits

The full paper combines formal models, worked designs and prototype measurements on synthetic credentials. Its reported crossover values depend on the chosen verification schedule, retained data, implementation and interpretation checks. The arithmetic and internal consistency have been reviewed; the original benchmark code and raw run archive are not included, so the publication package does not independently reproduce those measurements. The supplied cryptographic example tests are a separate reproducible result.

Attestation-volume ratios are scenario assumptions rather than universal requirements. The battery example assumes one qualified credential per 1,000–10,000 ordinary credentials. The one QEAA per five ordinary EAAs ratio for higher-risk settings is an author-provided Spherity estimate; the supplied package contains no supporting dataset or methodology. Neither ratio establishes decision assurance.

The paper does not claim that RDF is the only possible representation, that every read requires fresh verification, or that a graph alone proves safety or compliance. Source-backed and mixed-format architectures remain valid when their mappings and verification rules are governed. The author is chief executive and founder of Spherity GmbH, which funded the work and develops credential and organisational-identity software. The full paper contains the complete declarations and reference list.

Publication and selected primary references

  1. Stöcker, C. (2026). Proofs over Statements in Verifiable Evidence Graphs for the European Business Wallet, version 1.3. Spherity GmbH, 9 November 2026. The full paper contains 130 references.
  2. Stöcker, C. (2026). Why business wallet evidence graphs need a qualified W3C VCDM profile — executive brief. Spherity GmbH.
  3. W3C. Verifiable Credentials Data Model v2.0 and Verifiable Credential Data Integrity 1.0, Recommendations, 15 May 2025.
  4. W3C. Data Integrity EdDSA Cryptosuites v1.0, Recommendation, 15 May 2025.
  5. IETF. RFC 9901: Selective Disclosure for JSON Web Tokens, November 2025.
  6. European Commission. Proposal for a Regulation on European Business Wallets, COM(2025) 838.
  7. European Union. Regulation (EU) 2024/1183 amending the eIDAS framework.
  8. IETF. RFC 9955: Hybrid Signature Spectrums, July 2026, Informational.
Creative Commons Attribution 4.0 International

Open research

License and citation

The research content in this overview, both PDFs and the original figures is licensed by its named author under the Creative Commons Attribution 4.0 International License (CC BY 4.0), except for the third-party material identified below. Reuse must credit every named author, link to this canonical version and the license, and indicate whether changes were made.

Third-party rights: Third-party contexts and software dependencies retain the notices supplied in the publication package. The paper's CC BY 4.0 notice does not automatically assign a software licence to the verification scripts. Spherity names and logos are excluded from the research licence.

How to cite this work

Dr. Carsten Stöcker (2026-11-09). “Proofs over Statements in Verifiable Evidence Graphs for the European Business Wallet.” Spherity GmbH. https://spherity.github.io/spherity-research/proofs-over-statements-evidence-graphs.html. Licensed CC BY 4.0.

Direct answers

Questions this research answers

Why does this paper recommend JSON-LD credentials with RDF-based Data Integrity for industrial evidence graphs?

For retained evidence that many organisations repeatedly verify as graph data, a common W3C VCDM 2.0 JSON-LD profile with RDF-based Data Integrity aligns graph processing with the issuer-authenticated proof domain. It still requires governed vocabularies and contexts, compatible validation rules, source-snapshot recovery, issuer authority and status checks. JSON-LD syntax alone does not supply those guarantees.

Is a nested credential the same as a graph-proof credential?

No. Envelope-only B1 authenticates protected token inputs and disclosure commitments, then needs a governed binding to the graph interpretation. A nested credential that retains an inner Data Integrity proof can also authenticate that inner graph. When policy relies on both layers, both must be verified. A presentation-only wrapper need not be rechecked on every later use of independently verified inner evidence.

Are graph proofs always faster or more secure than SD-JWT VC?

No. The paper's cost advantage applies to specified repeated-verification paths with fresh envelope-to-graph expansion on synthetic fixtures. An envelope-only check under a trusted cached mapping can cost less, and a separate presentation test favours the tested SD-JWT implementation. Security depends on the complete profile, including interpretation, source binding, resource limits and application policy.

Does every edge in an industrial evidence graph need a QEAA?

No. Ordinary electronic attestations of attributes can support product, process, event and evaluation evidence where policy accepts them. Selected identity, mandate or other assurance-critical relationships may require qualified attestations under the applicable scheme. Qualification does not automatically propagate to neighbouring facts, derived conclusions or AI outputs.

What happens when a required qualified attestation is missing?

A justification path that requires that attestation cannot satisfy the target assurance level. Another admissible path may still satisfy the policy. If no admissible alternative exists, the decision cannot be justified at the required level; independent evidence elsewhere in the graph remains verifiable. A common credential format helps express and check dependencies but cannot repair missing assurance.

Are the plain and nested credential examples cryptographically validated?

Yes, under an explicit local demonstration profile. The plain credential has an Ed25519 Data Integrity proof; the nested version adds an ES256 SD-JWT signature and a disclosure commitment while retaining the inner proof. The main runner passed four acceptance controls and 13 expected rejection cases. A direct Node/OpenSSL audit passed nine assertions and eight negative checks, while sharing the RDF processing dependencies. These checks do not establish QEAA status or production conformance.

Does Data Integrity already support multiple signatures and post-quantum migration?

Data Integrity supports multiple proofs through proof sets and chains. A traditional/post-quantum hybrid still needs validated cryptosuites, key and credential bindings, mandatory-component verification and downgrade protection. Multiple proofs alone do not establish hybrid security, post-quantum assurance or qualified status. The paper's benchmarks do not measure PQC overhead.

How does this research support Trusted AI, industrial AI and physical AI?

Knowledge graphs express meaning, verifiable evidence graphs preserve authenticated sources and versions, and assurance graphs connect system claims to requirements, evidence and reasoning. Together they can help AI systems inspect the basis and authority for a proposed action. They complement safety engineering, monitoring, runtime controls and human oversight; cryptographic integrity does not prove factual truth or operational safety.