Research library
Papers are grouped by what they are for: why the problem matters, how the runtime solves it, what a governed scenario produced, what was measured, and what was measured earlier. Older benchmark runs are preserved unchanged and labelled with the hardware and date that produced them, so the library reads as a progression rather than a set of contradictions.
Looking for practical guidance?
Research remains the dated system of record for whitepapers, scenario evidence, and benchmarks. The StreamKernel blog answers focused architecture and governance questions, then links back here for proof.
Featured HTML whitepaper
A benchmark-backed financial-services whitepaper on screening, scoring, and provenance-stamping payment events with in-process ONNX inference inside the JVM, showing 20,480 records, 0.049ms average inference, and zero drops.
How this library is organised
Start with why the problem exists, then how it is solved, then what a governed run actually produced, then what was measured. The archive at the end holds earlier benchmark runs, preserved exactly as measured.
Why execution-time governance is becoming an infrastructure problem rather than a policy document problem.
A market thesis for operational AI: governed, indexed, repeatable, and closer to infrastructure than a conversational surface.
No performance claims
A visual explanation of disconnected model execution, local evidence, and single-JAR deployment for networks that cannot call out.
No performance claims
An infrastructure essay on unnecessary model-serving hops, energy cost, operational sprawl, and why local inference can matter.
No performance claims
How the governed execution boundary is built, and how it sits beside transport, model, and storage layers instead of replacing them.
An architects-and-engineers whitepaper on StreamKernel's embedded MCP control plane: a local, auditable, schema-described interface for agents to inspect runtime state, validate configuration, query evidence, and preflight guarded operations.
Historical · update note added Editorial note added on audit-chain scope: tamper-evident when enabled, per-runtime-instance, not an externally notarized ledger.
A transport-agnostic migration paper for teams moving between Pulsar and Kafka while preserving policy, transform, sink, and evidence behavior.
Historical · see benchmark registry PDF artifact. Migration reasoning remains valid; any throughput figures predate the Apple M5 Pro campaign.
A technical paper on generating embeddings inside the JVM runtime and delivering vector-ready records into MongoDB Atlas.
Historical · see benchmark registry PDF artifact. Embedding-pipeline architecture remains valid; throughput figures predate the Apple M5 Pro campaign.
A compact note on bootstrapping governed model artifacts and writing enriched records into Delta Lake from the same runtime boundary.
No performance claims
An architectural article on keeping scoring, policy, provenance, and delivery inside the event path instead of distributing them across services.
No performance claims
What a governed scenario actually looks like end to end, with the decision records and provenance a completed run produced.
A benchmark-backed financial-services whitepaper on screening, scoring, and provenance-stamping payment events with in-process ONNX inference inside the JVM, showing 20,480 records, 0.049ms average inference, and zero drops.
Historical · update note added Benchmark summary relabelled by hardware and status in August 2026. The 3.22M events/sec governed Kafka figure is corrected in place: it was not reproduced by the current build and is no longer published as a current result.
A public sector AI whitepaper on continuous inference provenance, SHA-256 audit fingerprints, FedRAMP-oriented evidence, mTLS Kafka delivery, OIDC token exchange, fail-closed OPA policy, authenticated observability, and retention governance.
Historical · update note added Editorial note added clarifying that the audit hash chain is per-runtime-instance and that nothing here constitutes FedRAMP authorization or agency accreditation.
A rich evidence whitepaper on normalizing, enriching, governing, routing, and auditing agent tool-call events before they affect enterprise systems.
Historical · update note added Editorial note added distinguishing this scenario run from claim-grade throughput campaigns, and clarifying audit-chain scope.
A defense telemetry whitepaper on evaluating and routing operational events in-process, with guardrail decisions, route labels, sensitivity scoring, reason codes, and provenance carried on every record.
Historical · update note added Editorial note added. Air-gapped deployment is an architectural property, not SCIF approval, NIPRNet authorization, or an authority to operate.
A healthcare pipeline whitepaper on detecting PHI fields, applying redaction or tokenization, running local AI enrichment, and writing governed Kafka records inside the same JVM process.
Historical · update note added Editorial note added. PHI handling supports a covered entity's compliance program; it is not a HIPAA certification.
A deeper financial-services paper on scoring payment events inside the runtime with in-process inference, policy decisions, audit output, and governed delivery.
Historical · see benchmark registry PDF artifact with no source template in this repository, so it could not be edited in place. Read its performance figures against the current benchmark registry.
A buyer-facing view of transaction scoring, AML control points, policy evidence, and in-process inference for regulated payment flows.
Historical · see benchmark registry PDF artifact with no source template in this repository. Read its performance figures against the current benchmark registry.
Measured runtime performance, with the measurement boundaries and comparison caveats stated rather than dropped.
A technical comparison explaining why StreamKernel's single-JVM, in-process AI runtime is positioned beside, not against, Flink and Spark clusters, with May 27 EOS benchmark evidence and external stream-processing reference points.
Historical · update note added Preserved as a May 27, 2026 Intel run. An update note states the current Apple M5 Pro exactly-once distribution and makes explicit that a single-node laptop result does not establish a throughput ranking against Flink or Spark.
Earlier runs preserved as dated evidence. They are kept unchanged because a benchmark documents what happened on a specific machine, build, and date.
A more narrative benchmark article connecting local evidence, throughput ceilings, caveats, and buyer-relevant proof paths.
Historical · see benchmark registry PDF artifact preserved as dated evidence. Superseded for headline purposes by the Apple M5 Pro campaign.
A historical Intel i9 proof note for the Kafka-sink NOOP producer ceiling (acks=1), distinct from the in-JVM DEVNULL ceiling.
Historical · see benchmark registry The original Kafka-sink NOOP producer-ceiling proof note. Preserved as Intel provenance; no current claim-grade Apple M5 Pro producer-ceiling replacement exists.
A run story tracing at-least-once optimization, payload behavior, and the mechanics behind a reproducible Kafka profile.
Historical · see benchmark registry Preserved as measured. The current at-least-once result is 2.201M records/sec on an Apple M5 Pro.
A Kafka delivery story focused on exactly-once semantics, reproducible settings, and throughput evidence under stronger guarantees.
Historical · see benchmark registry Preserved as measured. The current exactly-once distribution is 1.947M-2.312M records/sec on an Apple M5 Pro.
A long-run benchmark narrative for high-volume Kafka processing and the operational evidence around sustained record movement.
Historical · see benchmark registry The 563M-record run behind the historical 956K ops/sec Kafka-sink figure. Preserved as Intel provenance.
A MongoDB sink story for sustained insert throughput, operational settings, and the evidence behind a 95.5M document baseline.
Historical · see benchmark registry Preserved as measured. A later single Apple M5 Pro proof run measured 366,543 docs/sec at write concern w:1.
A security-path benchmark story for TLSv1.3, OPA-backed policy checks, deny behavior, and evidence under governed delivery.
Historical · see benchmark registry The historical 366K ops/sec mTLS + OPA path over 217M records. It predates the Keycloak OIDC profile and is not directly comparable to the current governed Apple M5 Pro path.
Papers marked “see benchmark registry” are PDFs. There is no source template for them in this repository, so they could not be annotated in place; read their performance figures against the current benchmark registry, which states the hardware, date, run count, and measurement boundary for every claim.
Commercial path
Use the commercial path for benchmark review, protected AI builds, plugin guidance, and regulated deployment planning.
