Our working log: real defects, rework and judgement calls from projects. Serialized alongside the WeChat account 「智能体驻场日记」.
A verification toolchain had run for a long time reporting “no issues found”. Only after building a defective sample did we find that two rules had never matched a single line since the day they were written.
Includes 8 checks any team can use: does every rule have a sample that must trigger it, is the warning volume sane, does a tool fail loudly or silently…
Full text on this site · also serialised on WeChat 「智能体驻场日记」
→A project that was “coded and statically clean” was judged not passed by our own gate — because stage 3 had no compile log: missing evidence counts as failure.
And how we handle unconfirmed site conditions: logged as a conditional pass — work continues, but the conclusion cannot serve as final acceptance.
Full text: WeChat account 「智能体驻场日记」
Many plants pay for a program that runs — and never receive what they need to maintain it themselves. Here are 6 deliverables worth writing into the contract, and what goes wrong when each is missing.
Plus one honest statement: static analysis cannot see semantic errors — never let "it passed static analysis" imply "the logic is correct".
Full text on this site
→