Make a retry failure easier to explain. Keep the test for your project, or give someone else a precise question to work on.
Runs in your browserNo account neededNothing uploaded
010203less noise. same failure.
Start with a supported trace.
This bench compares four named faults against our fixed Retry Proof v1 model. Use 1–12 synthetic operations with keys and values A/B. For your own implementation, run the exported test with a trusted local adapter.
Only ticks, A/B keys, A/B values, and received/lost delivery belong here. Keep real requests and credentials out of the trace.
Choose an example or load a supported trace.
02 / THE USEFUL PART
Find what matters.
⌁
Your comparison will appear here.
The result is computed locally. It is not a public contribution.
Each observation shows status / receipt / cumulative writes. “lost” means the caller received no response.
Inspect complete computed evidence
03 / TAKE IT SOMEWHERE
A test worth keeping. A question worth sharing.
Keep the useful output even if you never register. Publishing remains a separate choice.
01
Use it in your project.
A standalone Node test includes the selected trace and reference-derived expectations.
Run the test locally
node --test retry-regression.test.mjs
The default exercises bundled models only. Set RETRY_TRACE_ADAPTER to your trusted local adapter to test your own code. Adapter execution is bounded to three seconds, including import; it runs with your local permissions.
Save the candidate, then load it in the contribution workbench. Your existing profile, exact checks, and saved-operation confirmation control publication.
Continue checks relevant public changes when you request it. It does not schedule or wake an agent.
Download did not start? Save this file manually.
Already working in code?
The same bounded reducer runs locally in Node. Download the portable tools and one-run command, then follow the guide. No package install or model call is needed.
For a public result that already has a parent, use its exact handoff. A shortest subsequence of this input is not a proof of a globally shortest failure, vendor correctness, or concurrency safety.