Inference and agents, in the KingdomPay per token · Saudi RiyalDCP-Agent for Saudi business · agents.dcp.saAgents can rent a GPU · npx -y github:dhnpmp-tech/dcp-mcpEarn Riyal from your GPUPDPL · Saudi data residencyInference and agents, in the KingdomPay per token · Saudi RiyalDCP-Agent for Saudi business · agents.dcp.saAgents can rent a GPU · npx -y github:dhnpmp-tech/dcp-mcpEarn Riyal from your GPUPDPL · Saudi data residency
Benchmarks and evals · evidence firstSaudi workload claims stay gated
Benchmarks for Saudi AI workloads, locked to reproducible evidence.قياسات لأحمال الذكاء السعودي، مقفلة على دليل قابل للتكرار.
DCP exposes model benchmark metadata and provider benchmark contracts, but public Arabic-quality claims stay off until the dataset, harness, artifact checksum, and baseline report are all present.
Only models with a live verified provider and measured_at timestamp can show latency, cost, or quality numbers.
Source/api/models/benchmarks
arabic_quality_benchmarks
Arabic quality claims gated
Saudi customer-support and Arabic task claims wait for approved datasets, fixed harnesses, checksums, and baseline metadata.
Sourceclaim guard
provider_benchmark_contract
Provider benchmark contract
Provider benchmark records support fleet quality work, while public score claims still require GPU-host proof.
Source/api/benchmark
customer_evaluator_jobs
Eval job metadata records
Renter-scoped create/list/read endpoints now store draft eval metadata with dataset checksums and metrics; workers, results, billing, reports, and rankings remain off.
Source/api/evals/jobs
fig. 05 - benchmark feed to claim guard
§ 02 · API contractMetrics before marketing
Publish measured rows. Block quality claims until the run is reproducible.انشر الصفوف المقاسة. واحمِ ادعاءات الجودة حتى يصبح التشغيل قابلاً للتكرار.
The readiness endpoint is the contract agents can inspect before they quote latency, quality, or comparison claims. It is intentionally conservative: live latency can be visible while public Arabic-quality ranking remains false.
Seeded benchmark profile rows never become public numbers without live provider measurement.
Arabic task comparisons need approved datasets, harness version, baseline policy, and artifact checksum.
Eval job create/list/read endpoints are live for metadata records; workers, result APIs, billing, reports, and rankings remain blocked.
The worker readiness contract proves queue dispatch, result writing, and billing hooks stay disabled before eval execution.
The worker dry-run fixture proof simulates a draft eval queue item and writes only a temporary manifest, without changing job status or billing.
The result manifest schema defines checksum proof for future artifacts while result download APIs remain off.
The artifact storage policy defines renter/job-scoped result keys and checksum guards before object-store writes or signed downloads exist.
The result access policy requires renter ownership, result availability, artifact scope, and checksums before any result endpoint or signed download can go live.
The result endpoint route is renter-scoped but deliberately returns a disabled contract until signed-download proof exists.
The signed-download policy defines access, artifact, checksum, content-type, and expiry requirements without exposing a signed URL or object-store key.
The writer dry-run can create a validated manifest in temporary proof storage without production artifact writes.
Current contract state
A benchmark page is useful only if it refuses weak evidence.
This rail lets DCP look like a mature inference platform without drifting into unsupported claims. The page can show the benchmark contract now, then light up eval jobs and public reports only when the backend evidence says they are ready.
fig. 05 - benchmark feed to claim guard