The Unplug Test

An open acceptance specification for outage-resilient compliance evidence

Version 1.0, July 2026. Vendor-neutral. Anyone may run this test against any monitoring or evidence system. License: CC BY 4.0. Maintained at verityedge.eu/spec/unplug-test (versioned; changes are logged).

Why this test exists

Compliance evidence is only as good as its worst gap. Most monitoring tools produce dashboards and reports that assume the network stays up. Real edge sites lose WAN connectivity, and when they do, most tools silently drop the record. An auditor cannot distinguish "nothing happened" from "we were blind for two hours."

This specification defines a repeatable, falsifiable test for one question: does your evidence survive when your network does not?

Definitions

  • System under test (SUT): the agent or collector at the edge site plus the backend it reports to.
  • Outage window: a period during which the SUT's site has no outbound connectivity.
  • Complete backfill: after reconnection, every payload generated during the outage window is present in the backend, in order, with original timestamps, with no duplicates.
  • Continuity record: a tamper-evident log entry showing the outage window itself was recorded: start, end, buffered count, drain confirmation.
  • Silent gap: any interval in the evidence record that the system cannot itself prove was either covered or explicitly logged as an outage.

Test setup requirements

  1. One production-representative edge site running the SUT's standard agent, no test-only configuration.
  2. Payload cadence known in advance (for example one heartbeat per 30 seconds), so expected payload counts are computable: a 120-minute outage at 30-second cadence must yield 240 payloads.
  3. A method to cut WAN or block all outbound egress at the site (pull the uplink, firewall rule, or switch port down). Cutting power does not count; the site stays up, the network goes down.
  4. An independent observer able to record wall-clock start and end of the cut.

Procedure

Run all three windows, in any order, each as a separate test:

WindowCut durationExpected result
W130 minutesFull buffer, full backfill
W260 minutesFull buffer, full backfill
W3120 minutesFull buffer, full backfill

For each window:

  1. Confirm steady state: agent reporting at cadence for at least 60 minutes before the cut.
  2. Cut connectivity. Record wall-clock time.
  3. Wait the full window. Do not touch the agent or site.
  4. Restore connectivity. Record wall-clock time.
  5. Observe drain to completion without manual intervention.

Pass criteria (all must hold, per window)

  • P1 Count: backend payload count for the outage window equals cadence times duration, exactly.
  • P2 Order and timestamps: backfilled payloads carry original capture timestamps, not upload timestamps, and sort correctly against pre- and post-outage data.
  • P3 No duplicates: re-delivery does not create duplicate records (idempotent ingest).
  • P4 Continuity record: the outage window itself appears in a tamper-evident log: start, end, buffered count, drain confirmation. The absence of data is itself evidenced.
  • P5 No intervention: recovery required zero human action at site or backend.
  • P6 Reproducible export: an evidence export spanning the outage window can be generated two times with an identical cryptographic hash for identical scope and time range.
  • P7 Tamper check: an attempted modification of one historical record in the continuity log is detectable (for example a broken hash chain). If history can be silently edited, the test fails regardless of P1 through P6.

Reporting format

A published Unplug Test result must state: SUT name and version, cadence, each window's cut and restore wall-clock times, expected versus observed payload counts, the continuity record excerpt, export hash values from both export runs, and the name of the independent observer. Publish failures as well as passes; a spec without published failures is marketing.

What this test does not cover

  • Outages longer than the SUT's local buffer capacity. State your buffer ceiling explicitly (for example "at least 24 hours at standard cadence"); beyond it, full fidelity cannot be guaranteed by any system and claims otherwise should be treated with suspicion.
  • Power loss at the site, hardware destruction, or clock manipulation at the OS level.
  • The quality of the collected controls themselves. This test proves continuity of evidence, not sufficiency of controls.

Where Verity Edge stands today

We publish this specification, so we publish our own gaps against it. We make no claim that Verity Edge meets this specification yet.

  • P2, original capture timestamps: not met yet. The agent does not send a capture time today, so backfilled readings are stamped when they arrive.
  • P3, no duplicates through idempotent ingest: not met yet. If a delivery is repeated after a crash, the backend does not yet drop the repeat.

The planned fix for both is a capture time and a sequence number on every reading, and a backend that stores each reading once. We will update this note when both are met and tested.

Versioning

Substantive changes increment the version and are listed here. v1.0: initial publication, July 2026.

Download

Unplug_Test_Open_Spec_v1.0.md (Markdown, License: CC BY 4.0)

SHA-256: 6b7632cb08658a45664e7f73bda5624409cfbc5773fe75ab38929bb90fbcda48

Check it: sha256sum Unplug_Test_Open_Spec_v1.0.md