Validation Protocol

A method for testing
runtime claims.

This chapter defines a repeatable validation method for a specific product version, host, workload, and configuration. It does not report a successful test campaign.

On this page
Evidence status

No completed run evidence is published here.

The checked-in validation material does not contain a successful, request-bearing run. The only checked-in result is a failed smoke attempt that recorded zero requests. It cannot establish throughput, availability, recovery, enforcement, or release behavior.

The sections below are a protocol. Results become evidence only when a run records its version, configuration, timestamps, workload, traffic, faults, raw outputs, and acceptance decisions.

01 / Pre-register the run

Fix the test boundary before observing results.

Build identity

Record the management-plane build, launcher and agent release identities, and the exact workload package hash.

Environment

Record host resources, operating system, network path, runtime versions, and relevant external dependencies.

Runtime contract

Archive the validated infraveil.json, approval mode, gateway policy, restart limits, health definitions, and persistence paths.

Acceptance rules

Define pass, fail, and inconclusive thresholds before running traffic or injecting faults.

02 / Exercise the control path

Test identity, approval, execution, and outcome separately.

  1. 1. Enroll through the Trust Console.

    Confirm the customer key remains on the customer machine and the host receives only the scoped identity needed for its workspace.

  2. 2. Validate the runtime contract.

    Reject malformed or ambiguous infraveil.json input and record the accepted manifest revision.

  3. 3. Exercise release approval.

    For manual and allowlist modes, show that an unapproved exact hash stays pending. If auto is tested, record that it removes the per-release veto while enabled.

  4. 4. Correlate execution.

    Tie the approved hash to downloaded bytes, the launched process, health observations, route state, and the resulting receipt.

03 / Exercise the data path

Measure only traffic that reaches the configured gateway.

Drive a recorded mix of successful requests, application errors, malformed requests, slow bodies, streaming responses, and WebSocket traffic through managed routes. Compare client results with gateway decisions, upstream application logs, and counters.

Report request rate, latency distribution, response codes, dropped connections, resource use, and any missing telemetry. A test at the origin does not establish protection against upstream volumetric DDoS, traffic that bypasses the configured gateway is outside the gateway result, and provider-layer mitigation must be validated separately.

04 / Inject bounded failures

Observe stop conditions as closely as recovery.

Control-plane outage

Interrupt synchronization; record workload continuity, stale-state labeling, queued activity, and reconnect ordering.

Process failure

Crash a supervised service until its configured restart budget is exhausted; verify that the runtime stops thrashing and reports the failure.

Bad candidate

Deploy an unhealthy but valid release and observe the configured health gate, stop, or rollback behavior.

Disk pressure

Constrain workspace or cache storage; record whether writes stop safely and what remains recoverable.

Dependency failure

Remove a required upstream service and distinguish running, healthy, reachable, and current states.

Revocation while offline

Record the delay between hosted revocation and local application after the host reconnects.

05 / Inspect custody and persistence

Record what is stored, not what the mode name implies.

For source-custody mode, enumerate application files, the manifest, package, cache, and configured persistent paths on disk. For no-source attach, verify that a fresh package contains the manifest only and separately inspect whether files from any earlier hosted package remain in the workspace.

Repeat the inspection after restart, failed deployment, rollback, cleanup, and offboarding. Record permissions and deletion behavior for each path.

06 / Publish the record

A result needs enough context to be reproduced.

RecordRequired contentsWhat it can support
Run manifestVersions, hashes, host, topology, configuration, start and end timeReproducibility boundary
Traffic recordGenerator settings, completed requests, response distribution, latency, errorsObserved behavior for that workload
Fault timelineInjected event, expected response, actual timestamps, stop or recovery stateObserved failure handling
Evidence bundleRaw outputs, receipts, logs, counters, gaps, and reviewer decisionAudit and independent review

Results apply only to the recorded build and environment. A later runtime, policy, workload, or infrastructure change requires a new run or an explicit explanation of why the prior result still applies.