DraftSystems / DBEDBT short / demo or ESWC in-use/resources — results-free systems description; research-track submission deferred until the specified evaluation is executed
Filter-as-Query: Filtered Approximate Nearest-Neighbour Search over SPARQL, where the Filter is an Exact BGP over the Engine's Own Dictionary Ids
An RDF-native filtered-ANN integration: the filter on a vector-neighbour variable is the join-connected sub-BGP of the query itself, evaluated exactly by the engine and materialised as an id-mask over the shared dictionary-id space (no metadata mirroring, no boundary id translation). The pre≡post answer contract is enforced as machine-checked invariants and holds unconditionally on the path the SPARQL surface reaches, which is answer-exact — both physical strategies rank the complete admitted pool. The crate's approximate filtered traversal is a separate path the paper claims nothing for. Positioning is explicit that the constraint pushdown is a static instantiation of sideways-information-passing (magic sets, semi-joins, RDF-3X SIP) rather than a new strategy; related work covers the 2023–26 filtered-ANN wave incl. the engine-integrated systems (VBASE/NaviX) it is closest to. The performance evaluation is specified (baselines, workloads, falsification criteria) but unexecuted, pending the canonical runner.
Evidence: Deterministic only: recall floors (0.95 unfiltered, 0.90 approximate-filtered — sanity checks on a traversal the SPARQL surface does not reach, labelled as such) + asserted pre-filter ≡ post-filter equivalence (single / transitive / cyclic), unconditional because the SPARQL-level filtered path is answer-exact, + the cost-model crossover constant. A canonical invariant record that went false now aborts the paper build rather than rendering blank. No latency/throughput claim; the performance evaluation is specified (not pre-registered — no registry deposit exists) and unexecuted, blocked on the canonical runner.
Publishable nowSemantic WebISWC resources / in-use track (or demo)
Library-Level Solid Access-Control Conformance: WAC and ACP Decision-Parity Ratchets Joining a Cross-Family Scoreboard
A library-level conformance signal for both Solid access-control models: a per-construct, fail-closed-checked decision-parity corpus for WAC and for ACP, each ratcheted to a monotone floor and registered as a row of the single cross-family scoreboard with a drift guard. The honest evidence is conformance, not speed, and no security property is claimed; HTTP / CTH wire conformance is explicitly out of scope.
Evidence: Deterministic conformance ratchet floors: Solid WAC decision parity (12) and Solid ACP decision parity (12), as the two newest rows of a 7-suite cross-family scoreboard totalling 3442, each floor CI-enforced, monotone, and guarded against drift. No latency claim and no security/soundness claim; library-level decision parity only, not HTTP/CTH wire conformance.
DraftSemantic WebESWC research track (policy) — draft; not submittable until the §5.3 comparative study runs
An ODRL Policy Bridge for SPARQL Access Control: Fail-Closed Compilation of Usage Policies into a Queryable Solid Access-Control View
Compile, don't co-evaluate: a matched ODRL Permission/Prohibition compiles into the same triples the engine's existing, queryable WAC/ACP view already understands — no second enforcement engine, and every compiled decision is auditable, provenance-tagged RDF (the paper walks one policy end-to-end: Turtle policy → compiled auth/provenance triples → SPARQL audit query and its result). The contribution is the fail-closed lifecycle discipline compilation demands: ODRL's prohibit conflict strategy realised structurally by allow-minus-deny set subtraction (the perm and spec-default invalid strategies are not representable — a disclosed, first-class limitation), asymmetric three-valued deny retraction on policy refresh (positioned against decision caching, authorization recycling, Zanzibar-style consistent authorization, and materialised-view maintenance — not only ODRL enforcers), per-session re-checked conditional grants with safe one-shot fallback, and atomic count budgets. Evidence, honestly: four machine-checked invariants are the only direct bridge evidence; the WAC/ACP decision-parity floors are context for the pre-existing target layer, not bridge evaluation; the specified comparative decision-agreement study vs ODRE/OAC has not run (hence draft — not submittable until it has). The federated ODRL→MPC / ODRL-Duty→ZK composition is deferred, unbuilt, and claims nothing (sq-qhy4).
Evidence: Honestly partitioned. Direct evidence (all of it, today): deterministic, test-proven answer-safety invariants of the bridge (prohibit-strategy set subtraction correct through unchanged enforcement; asymmetric fail-closed deny retraction; recipient constraints persist as re-checked conditional grants with one-shot fallback; atomic stateful count enforcement). Context — NOT bridge evidence: the CI-enforced, drift-guarded Solid WAC (12) and ACP (12) decision-parity ratchet floors of the pre-existing target layer. PENDING — the blocking gap to submission: a fully specified (not pre-registered: no registry deposit exists) decision-agreement study vs the ODRE enforcement engines and an OAC-style matcher over the cited systems' own policy corpora. No latency claim; no novel-semantics claim; no in-use claim; the federated/ZK disclosure half is explicitly deferred (research-grade crypto, not externally audited; sq-qhy4).
WIP · arXiv onlyCrypto (WIP)PoPETs / a security workshop (negative-result · security-engineering lessons)
Cannot Certify, So We Encode the Obligation: A Collaborative-zk-SNARK Re-Audit as a Witness-Validation Negative Result for Federated SPARQL
An adversarial re-audit of an engine's intended collaborative (multi-prover) zk-SNARK path against the CRYPTO'25 (eprint 2025/1026) failure modes. The honest disposition is a negative result: the path is unbuilt, so soundness cannot be certified and every lens is RE-OPEN. The durable output is R-WV, a witness-validation-before-proving test obligation encoded as a build-time gate. NO security, privacy, or attestation property is claimed.
Evidence: Negative result — asserts NO proven security/privacy/soundness/attestation property (the collaborative path is unbuilt: 6 proof/attestation entry points fail closed with NotYetImplemented). Committed structural counts only: 4 re-audit lenses, all RE-OPEN; a 5-clause R-WV witness-validation obligation encoded as a build-time gate; 12 prior single-prover findings under the open external-audit gate. Estate is research-grade and not externally audited; cites the gates sq-qhy4 (external single-prover audit, open) + sq-9hrn (coZK re-audit). No performance claim.
DraftCrypto (WIP)PoPETs / a security or semantic-web privacy workshop — WIP; C-family, so arXiv-only until the external audit gate (sq-qhy4) closes and a canonical proving-time runner replaces the gate-count cost model
A Single-Prover Zero-Knowledge Architecture for Verifiable SPARQL over Committed RDF Graphs
The system design of a single-prover zero-knowledge stack that proves a SPARQL result is a genuine evaluation over committed, issuer-attested RDF graphs — without disclosing the graphs. It describes the pieces that make the prover-as-adversary model tractable: per-graph Poseidon2 commitments over the RDFC-1.0 canonical form bound to Schnorr issuer attestations; a fixed, named family of 13 circuit kinds (31 compiled members) each proving one operator instance of a monotone, open-world-conforming SPARQL fragment (BGP scans, datatype-bucketed value FILTER, a hidden-credential equality JOIN); a JSON manifest that composes sub-proofs through checkable binding edges; and a verifier that re-derives the circuit identity and the claimed statement from the query text and the relying party's trust anchors — a 12-obligation, 4-audit-gate fail-closed pipeline that trusts nothing the prover declares. Cost is reported as deterministic bb UltraHonk gate counts from the regression-gated snapshot plus a linear proving-cost model over them (no wall-clock headline: development timings are non-canonical, so the missing per-host constant is named rather than quoted). Security is stated under the OPEN external-audit gate: the verifier is internally re-audited (12 findings found and remediated, each pinned closed by a standing forge-negative regression test) but NOT externally audited, so no proven property is claimed and the internal audit is framed as necessary-not-sufficient.
Evidence: Deterministic artifact facts only, asserting NO proven security/soundness/privacy/attestation property (external audit gate sq-qhy4 OPEN). Canonical headline evidence: bb UltraHonk gate counts from the regression-gated snapshot (scan lattice 5991–34821; composable filter lanes gate-identical at 17416; join 7025–18681; revocation/hidden-issuer/holder members) + structural verifier facts (13 circuit kinds, 31 compiled members, 12 fail-closed binding obligations, 4 audit gates) + the internal-audit posture (12 confirmed findings, each pinned by a 1:1 forge-negative regression test). NO wall-clock headline — a linear proving-cost model over the gate counts is published because development timings are non-canonical (no canonical runner for this family). Cites sq-qhy4 throughout; hidden-holder tiers explicitly not-yet-sound; dual-leaf lane an accepted invariant downgrade.
DraftSemantic WebISWC / ESWC research (empirical) after the scale-up (to be pre-registered); K-CAP or an LLM+KG workshop as a pilot
Formal Ontologies for LLM-Agent Knowledge Management: schema.org vs gUFO vs DOLCE
A controlled pilot asking whether typing a project knowledge graph under a formal upper ontology helps an LLM agent answer knowledge-management questions — and which ontology. Holding instance data, agent, tasks, and grading fixed across four committed typing overlays in a fully-crossed single run, two overlays beat the untyped baseline — schema.org-as-top by the largest margin, DOLCE modestly — while gUFO fell below it. The overlay ranking is consistent with LLM training-data fluency, stated as a correlational hypothesis with its confounds named (overlay verbosity/closure noise, hand-authored overlay quality), not a demonstrated mechanism. Every accuracy figure is an indicative single-run measurement from one small agent model (Claude Haiku); a multi-run scale-up, to be pre-registered before it runs, is the stated gate before any venue submission.
Evidence: Indicative-only pilot: deterministically-graded answer accuracy of a non-deterministic LLM agent (Claude Haiku, one fresh instance per condition-task pair) over a committed 16-task corpus fully crossed with 4 committed ontology-overlay conditions (bench/fo-km). The only canonical records are structural corpus counts (16 tasks, 4 conditions); every accuracy/abstention figure is environment=indicative (single run, heuristic gold-key grading, dev work-box) and is structurally barred from headline citation. No latency claim, no significance claim.
WIP · arXiv onlyCrypto (WIP)PoPETs / IEEE S&P SoK track — WIP; C-family, so arXiv-only until the external audit gate (sq-qhy4) and the frontier's publication-absence claims are re-verified at submission time
SoK: Systematizing Verifiable and Private Federated SPARQL
A systematization of verifiable and confidential federated SPARQL under a stated method (inclusion criteria plus a dated search protocol) along three independent axes — cryptographic mechanism (single-prover ZK over committed graphs / MPC evaluation over secret-shared graphs / attested-input binding), prover topology (one holder vs N mutually distrusting holders), and adversary model (semi-honest / covert / malicious, crossed with output guarantees under Cleve's bound) — axes assembled from established MPC and collaborative-proof taxonomy, applied jointly to federated SPARQL. Contributes an operator-level capability matrix cross-referenced to the published systems, exposing the disclosed/hidden two-regime split and the disclosed-key-join cost advantage global IRIs give RDF — analyzed together with its privacy price (the cleartext join disclosing exactly the cross-source linkage that source-unlinkability protects); a catalogue of settled negatives (Cleve, the post-quantum boundary of the signature and SNARK layers — with the hash-based Poseidon2 commitment binding stated as PQ-resilient, blank-node cross-graph joins, in-circuit entailment) distinguished from merely vacant cells; and a method-bounded statement of the open frontier — verifying issuer signatures over a secret-shared witness inside one source-unlinkable collaborative proof, a composition no publication in the searched peer-reviewed venues or preprint archives instantiates as of the stated survey dates. Anchored in one research-grade estate whose capability tiers are self-reported documentary provenance (a stated independence limitation): single-prover layer not externally audited (sq-qhy4, open), MPC layer semi-honest by default, collaborative-proof path unbuilt and fail-closed. No security property is claimed and no wall-clock number appears.
Evidence: Systematization — asserts NO proven security/privacy/soundness/attestation property and cites no wall-clock measurement. The only build-injected counts are the deterministic structural facts already gated as canonical for the C-family estate (fail-closed collaborative entry points, re-audit lenses all RE-OPEN, witness-validation obligation clauses, prior single-prover audit findings under the open gate sq-qhy4). Capability tiers are documentary (spec corpus + reconciled capability review), stated as such in the paper's limitations.
DraftSystems / DBISSTA 2027 (testing track) — first choice; PVLDB Vol 20 rolling or FSE 2027 — second choice. Draft; not submittable until the campaign (bead sq-gum8.11) yields confirmed third-party bugs
Reifying the Error: Metamorphic and Differential Logic-Bug Testing for SPARQL Engines
SPARQL has no dedicated logic-bug testing work. We re-derive TLP and NoREC for SPARQL by reifying its third evaluation outcome — a type error, not a value — with the language's only error-absorbing forms, yielding a partition that provably recomposes the unpartitioned query under the SPARQL 1.1 spec. The merged instrument (crates/sparq-metamorph) includes a TLP + NoREC + differential oracle suite whose self-tests assert non-vacuity against a seeded wrong-result mutant on the real engine. This is a first draft against the merged instrument only; the cross-engine bug-hunting campaign has not run, every campaign table is an explicit placeholder, and the paper's honest publishability condition — previously-unknown, developer-confirmed bugs in third-party engines — is not yet met.
Evidence: Deterministic instrument self-tests only (no campaign results yet): 3 oracle types (TLP, NoREC, differential) each asserting non-vacuity against a seeded wrong-result mutant on the real sparq engine; oracle correctness verified across 50 generated seeds on the pristine engine; grammar exclusion list verified across 200 seeds (no banned non-deterministic construct). Campaign-dependent evidence (bugs confirmed/reported/rejected) is pending bead sq-gum8.11 and is shown as an explicit PLACEHOLDER in the paper. No third-party bug is claimed.
DraftSystems / DBPVLDB Vol 20 rolling (monthly deadline through 2027-03-01); ICDE 2027 R2 (2026-11-11) or EDBT 2027 cycle 3 (2026-10-07) as alternatives. Draft; not submittable until the canonical-host evaluation (bead sq-vw3ax.12) yields the gated performance/memory results
One Substrate, Many Standards: An Out-of-Core SPARQL Engine and a Measured Zero-Overhead Evaluation Core Across the W3C/OGC Spec Families
RDF triple stores historically force a speed/breadth/frugality three-way trade-off. This paper describes an out-of-core engine reaching for all three: it stores triples memory-mapped in six permutations with inline-tagged ids, evaluates queries via mixed binary/worst-case-optimal/bind joins, and extracts its evaluation core into a single substrate shared unchanged across SPARQL, OWL profiles, RIF, stream processing, GeoSPARQL, and SHACL — validated as behaviour-neutral by deterministic layout ratchets and a cross-family conformance scoreboard. Competitive performance and memory claims are gated on a canonical-host evaluation (not yet run) and are not asserted.
Evidence: The architecture (six permutation indexes, inline-tagged ids, mixed join families), the substrate extraction (single leaf crate consumed by all standards), and the conformance breadth (OWL RL/EL/QL, RIF, RSP, GeoSPARQL, SHACL, each pinned with a ratchet floor) are real and traceable to crates/sparq-core, crates/sparq-engine, crates/sparq-substrate, crates/sparq-reason*, crates/sparq-rsp, crates/sparq-geo, crates/sparq-shacl. Competitive performance and memory figures (latency vs native QLever/Virtuoso, bytes-per-triple vs HDT/qEndpoint, substrate zero-overhead delta, Sparqloscope and qEndpoint comparisons) are deterministic gating measurements on bead sq-vw3ax.12 and carry environment=canonical; they are pending and not claimed here.