Question and method
Does a clean process restart preserve the fictional demo’s state, and can its retained public updates reproduce that state? Tested 8 November 2026 in Sydney on public commit 4271dfb831cbc8609548c88518340a76a9b56672, using Node 26.7.0 and pnpm 11.22.0 on macOS arm64.
The fixed seed was meridian-public-demo-v1, with 24 fictional residents. One fresh save advanced 60 canonical minutes continuously. Another advanced 30 minutes, closed cleanly, reopened in a new operating-system process, then advanced 30 more. The test compared full canonical snapshots and replayed all 60 public updates. Canonical minutes measure simulated time, not elapsed time.
Jake commissioned the check; a coding assistant prepared and ran it, and a second assistant independently repeated it. Population, seed and four opening events are preset by people; existing code rules drive resident behaviour. No model provider was configured. Instrumented network APIs recorded zero connection attempts.
Actual result
| Check | Observed result |
|---|---|
| State immediately after restart | Complete canonical and public snapshots matched the saved state. |
| Final state versus uninterrupted run | Identical canonical state at minute 60; public data matched apart from each fresh save’s independent stream ID. |
| Public replay | All 60 updates reproduced the final public snapshot. |
| Final record | Canonical sequence 1,866; public cursor 63; 86 retained task-memory entries. |
| Missing-update control | Skipping the first update in memory was rejected with public resync required. |
| Independent repeat | All seven checks passed again. No test failures. |
Both runs produced canonical snapshot SHA-256 48c55fe257ba53e9d70d582f5adcde6d553bb0dc266fcf9cb20bbd62c89de08e. This digest covers the canonical JSON snapshot, not the SQLite file.
Limits
This checks one fixed seed over 60 canonical minutes and a clean shutdown. It does not test crashes, power loss or long-term reliability. Public history retains 128 updates; complete indefinite replay is not established. Stored task-memory entries do not demonstrate learned competence. This local fictional demo is a separate release from the hosted city. The result establishes neither AGI, consciousness, autonomy nor self-replication.
Repeat the check
Use Node 26.7.0 and pnpm 11.22.0. In a new local folder, download the test harness alongside a checkout named public-demo. Keep the source unchanged. The test creates new save directories and preserves them.
git clone https://github.com/CreeleyDigitalSystems/meridian-public.git public-demo
git -C public-demo checkout --detach 4271dfb831cbc8609548c88518340a76a9b56672
cd public-demo
pnpm --filter '@meridian/world-service...' install --prod --frozen-lockfile --ignore-scripts
cd ..
node --import ./public-demo/scripts/source-loader.mjs ./replay-experiment.mjsThe command prints a summary and a new experiment-runs/run-… folder. Compare its final canonical digest with the value above. Fresh stream IDs and process IDs are expected to differ. The harness instruments network APIs; it is not an operating-system network sandbox.
Inspect the first summary, independent repeat summary, pinned source file hashes, and full recorded inputs, snapshots and updates: continuous run, first 30 minutes, resumed 30 minutes. For comparison with existing project checks, see the public demo tests.
Useful feedback
A reproduction with its exact version, runtime and output is useful feedback, including failures. Read how to contribute.