DN Runtime Recovery Local Baseline v0.1 — 21 October 2026

Scope: actual local LangGraph framework execution with deterministic synthetic
nodes and a separate local SQLite receiver. No LLM, managed cloud, wallet,
payment, production website change, or external service call in the harness.

Reproduce in a disposable local Python environment (tested on Python 3.12.14,
Linux x86_64, SQLite 3.53.1). Installing dependencies requires internet access;
executing the harness does not. Optional LangSmith tracing is disabled in code.

  python -m venv .venv
  .venv/bin/python -m pip install -r requirements-lock.txt
  .venv/bin/python recovery_test.py run --repeats 10 --output my-results.json

Use the published results-v0.1.json as the reference; output timestamps and
environment records will differ. Preserve your new result as a separate run.
Expected effects and acceptance criteria should match on an equivalent setup.
To check consistency of the original reference receipt and its source hash:

  python verify_results.py

The runner creates temporary SQLite databases, launches child processes, and
abruptly exits each first-phase worker with exit code 86. It removes its own
temporary trial directories after recording the receipt. Do not point it at an
existing database. Timeout or setup failures abort the run instead of quietly
dropping a trial. The published final suite includes every one of its 60 trials;
development smoke runs are not part of that final suite.

Configurations are deliberately chosen controls, not vendor defaults:
- Restart: in-memory vs SQLite checkpoint, same thread identity after restart.
- Uncertain outcome: receiver commits before worker acknowledgement; retry uses
  a new request key vs the same key. Both variants use SQLite checkpointing.
- Revocation: policy changed while paused; stale cached allow vs receiver-side
  recheck in the same local transaction as the effect. Both use SQLite.

An independent read of the receiver ledger counts committed synthetic effects.
For restart, exactly one effect and one preparation are required; for uncertain
outcome, exactly one effect; for revocation, no effect and a blocked outcome.
The worker result is also recorded. This is not a remote independent reviewer.

Limits: one host and framework version, serial deterministic cases and selected
injection points. No power failure, kernel crash, remote network partition,
concurrent revocation, model decision quality, latency benchmark, full-cost
comparison, or production failure-rate estimate. Receiver-side idempotency is a
local contract; an arbitrary external provider may not support it. A stable key
must also be bound to caller, intent/payload and retention rules in real systems;
this fixture does not test conflicting payloads or key expiry.

The Managed Runtime Dataset v0.1 remains an immutable documentation baseline.
This execution release is distinct and does not change Agent Trust scores or
imply any hosted provider passed these tests.

License: DN original source and test descriptions MIT (see LICENSE.txt);
DN original results CC BY 4.0. Dependencies retain their own licenses.
Article: https://decentralised.news/dn-runtime-recovery-test
Corrections: https://decentralised.news/contact
