Artifact identity
Model family, semantic version, immutable digest, tokenizer, precision, context configuration, and release state.
A reviewable path from
model artifact to evaluation access.
This publication defines the evidence record Software Spark expects for SPARK model programs and workload-specific deployments. It is technical process evidence—not a substitute for a measured benchmark result.
Read the publication ↓01 / EVIDENCE RECORDFour records before
Four records before
a release claim.
A deployment should be reconstructable from immutable identifiers, disclosed settings, retained results, and an approval decision.
Development disclosure
Permitted data categories, curation controls, post-training stage, evaluator lineage, and exclusions.
Serving profile
Runtime version, accelerator model and count, memory profile, batching, cache policy, concurrency, and region.
Evaluation record
Dataset version, metric implementation, raw output retention, reviewer decision, failure analysis, and promotion gate.
02 / RELEASE STATEPrivate Preview means
Private Preview means
evaluation—not production.
A Private Preview model card may describe intended use, limitations, target context, and evaluation plan. It does not imply general availability, public weights, a public SDK, a production SLA, or independently verified performance.
GATEPRIVATE PREVIEWEVALUATION ACCESSPRODUCTION
ArtifactVersioned targetScoped artifactImmutable approved digest
EvidenceEvaluation planCustomer test reportAcceptance record
InfrastructureProposed profileMeasured profileContracted profile
AccessInvite onlyScoped credentialProduction credential
CommitmentNone impliedEvaluation termsSigned service terms
03 / API EVIDENCE FORMATRecord the request,
Record the request,
response, and route.
Evaluation responses should include a request identifier, model version, route or deployment identifier, timing fields, usage, and the structured result. The exact endpoint and credential are issued only under evaluation access.
Download the JSON evidence template →{
"request_id": "req_<opaque>",
"model": "spark-atlas-1.1",
"release_state": "evaluation_access",
"deployment_profile": "<disclosed profile>",
"timing_ms": { "ttft": null, "total": null },
"usage": { "input_tokens": null, "output_tokens": null },
"result": { "status": "customer-measured" }
}04 / NEXT PUBLICATION
Measured results remain pending.
A public result will not be filled with estimated numbers. It will be published after a SPARK model artifact, GPU environment, benchmark dataset, concurrency profile, and raw results are available for reproduction.
Review benchmark protocol →