Curious Compute / Put it to work
Take the question
into your work.
Small checks to discuss with your team or adapt in a test environment. Each starts with a claim and the condition that changes it.
Proposed exercises, not tested product recommendations.
Ship an agent
Before an agent retries a refund
- The claim
- A timeout means the action failed.
- The missing condition
- The action may have finished even when its reply never arrived.
- Ask this
- Can your system distinguish a missing reply from a refund that never happened?
What to look for
- Keep the approved refund reference across recovery attempts.
- Use the provider’s supported recovery rules; never assume a fresh key is safe.
- Successful recovery produces one refund. An unresolved outcome remains unknown and paused.
Open the fictional test fixture
{
"fictional": true,
"kind": "fault-injection-plan",
"approvedRefund": {
"reference": "CC-REFUND-01",
"amount": 40,
"currency": "USD"
},
"cases": [
{
"failure": "request lost before execution",
"initialRefundCount": 0,
"observedByAgent": "timeout"
},
{
"failure": "reply lost after execution",
"initialRefundCount": 1,
"observedByAgent": "timeout"
},
{
"failure": "request still processing",
"initialRefundCount": "unknown",
"observedByAgent": "timeout"
}
],
"record": [
"persistent operation reference",
"attempts",
"provider state",
"actual refund count",
"recovery decision"
],
"edgeCases": [
"late reply",
"duplicate event notification",
"provider key no longer retained"
]
}
Download fixture (JSON)A test plan for a simulated payment service, not an executable payment integration or a reliability certification.
Evidence reviewed 2026-09-26 · Inspect the sources
Read the explanation →Evaluate an answer
Before you trust a quoted answer
- The claim
- A quotation from the right manual supports the answer.
- The missing condition
- The quotation may describe a different task, version or exception.
- Ask this
- Did search find the right passage, and did the AI receive enough context to use it?
What to look for
- Keep search selection, delivered context and answer correctness as separate checks.
- Answer 5 seconds for user-waiting requests; 30 seconds for background exports.
- If the applicable section is removed, acknowledge missing evidence instead of borrowing the other limit.
Open the fictional test fixture
{
"fictional": true,
"document": "Lyra manual",
"version": 2,
"sections": [
{
"heading": "Background exports",
"text": "The limit is 30 seconds."
},
{
"heading": "Requests a user waits for",
"text": "The limit is 5 seconds."
}
],
"questions": [
{
"question": "How long may a request keep the user waiting?",
"expectedSeconds": 5,
"requiredHeading": "Requests a user waits for"
},
{
"question": "What is the time limit for background exports?",
"expectedSeconds": 30,
"requiredHeading": "Background exports"
}
],
"variants": [
"without headings",
"with verified headings",
"applicable section removed"
],
"record": [
"retrieved passage",
"exact text delivered to model",
"correct section found",
"heading delivered",
"answer uses correct limit"
]
}
Download fixture (JSON)A fictional diagnostic fixture. No AI evaluation has been run; passing this example does not establish production reliability.
Evidence reviewed 2026-09-26 · Inspect the sources
Read the explanation →