Demo (E2E) Guide

A scripted, end-to-end walkthrough that demonstrates dispatch → execution → validation → aggregation → quarantine.

Scenario

Operators: matmul → ewise_add → unary_relu → reduce_sum
Sizes: 64×64 · 64×16 for fast laptop runs
Topology: 1 orchestrator, 2 agents (1 trusted, 1 untrusted)

Topology & Startup

Start orchestrator (docker-compose or binary)
Start a trusted agent (registered credentials)
Start an untrusted agent (default permissionless)

See also: Getting Started · Cloud Deployment

Job Submission

Submit a demo job with embedded inputs (or references) using CLI/curl. Capture job ID.

Expected orchestrator logs:

Task graph creation → dispatch events → validation receipts → final result

Validation & Quarantine Flow

matmul dispatched to untrusted agent → 2 rounds of Freivalds → accept on match
ewise_add on trusted agent → sampled index checks (still applied at low rate)
unary_relu on untrusted agent → sampled invariants
reduce_sum on trusted agent → partial recompute on slices

Failure injection:

Toggle fault in the untrusted agent for matmul → Freivalds mismatch → quarantine → reschedule to trusted agent → success

Expected Outputs

Final scalar result and checksum
Validation receipts per task (store/print JSON)

See also: Operator Test Vectors · Observability

Troubleshooting

Agent not visible: check capability advertisement and network settings
Validation fails intermittently: verify numeric tolerances and deterministic kernels
Quarantine loop: clear agent state or restart without fault injection

Next Steps

Automate this flow with a sample docker-compose.yml
Add dashboards (Prometheus/Grafana) from Observability

Variant: Using Implicit Data

Run the same scenario but pass inputs as references (URIs/handles) instead of embedding data:

Register datasets in the Data Service or supply file:///pre‑signed https:///s3:// links
Observe reduced payload sizes and that validation fetches only sampled slices

See: Data Sources & Data Service