Self-Verification: The End-to-End Gate, Premature Completion Claims and Automatic Intervention on Doom Loops
The two most expensive unattended failures are the premature completion claim and the doom loop. Today the harness gets an end-to-end gate and stall detection, and the emphasis falls on what happens after detection: switching tasks, rolling back, and stopping with an alarm each have their own test.
Today's Goals
- Can explain why the model saying it is finished is never grounds for flipping the state, and say what an end-to-end check should and should not verify
- Can state the trade-off behind running end-to-end checks at the HTTP layer instead of through a browser, and name the class of defects it cannot see
- Can describe the three interventions available once a stall is detected and the test behind each, and explain why detection without intervention is worth nothing in an unattended run
D1 through D4 left an obvious hole open: verification was only ever a shallow check, and a patch that landed in the file counted as a pass. Today closes it, and then deals with the new problems that closing it inevitably creates. When you finish, scroll back up and tick off the three goals.
Plain-Language Walkthrough
The building inspector does not grade the crew's own report
Halfway through a renovation you ask the foreman whether the plumbing is done. He says it is done.
Do you believe him?
You are not questioning his character. You simply know one thing: he is telling you that he finished what he understood the job to be, and what you want to know is whether the plumbing actually works. Between those two sentences stands an inspector, somebody who does not interview the crew and instead goes and turns the water on.
An agent is no different. When it reports that F07 is implemented, what it is telling you is that it wrote some code according to its own reading of the task. It has never run the service. On what grounds would it know that what it wrote works?
So the whole day rests on one plain sentence: the model saying it is finished is never grounds for flipping the state. The only thing that qualifies is an independent check that goes out into the environment and collects evidence.
Interview questions
An agent reports a task done - what evidence should your harness use to decide whether to believe it?Agent 报告任务完成,你的 harness 该用什么依据决定信不信?
Common in ChinaCommon overseasIntermediate#self-verification#false-completion#gates🔒 Unlock this chapter to see the answer keys.
Browser automation versus HTTP-only end-to-end verification - what does each buy and cost?端到端验证用浏览器自动化和只打 HTTP 接口,各有什么得失?
Common in ChinaCommon overseasIntermediate#end-to-end#trade-offs#capability-boundary🔒 Unlock this chapter to see the answer keys.
How do you detect an agent spinning in place, and what do you do once you have?你怎么发现一个 Agent 在原地打转?发现之后该做什么?
Common in ChinaCommon overseasDeep dive#stall-detection#intervention#rollback🔒 Unlock this chapter to see the answer keys.
A verifier built to catch false completions - how do you confirm it is not tautological itself?一个防谎报的验证器,你怎么确认它自己不是恒真的?
Common in ChinaCommon overseasDeep dive#assertions#mutation-testing#probe-design🔒 Unlock this chapter to see the answer keys.
