Is the AI-written code any good?

The failure mode is not bad syntax. A method returning a constant under a confident name, an async that never waits, an empty type, a dead branch, a catch that swallows the failure. It compiles, it type-checks, it passes the gate — and it does nothing. Watchdog reads this as the Generation-Debt lens: the quality signature of unreviewed code, judged on its own merits regardless of who or what produced it, reported with the rule that fired and the exact file and line rather than as an impression.

Each change is right where you read it and wrong once you see the whole: a fourth way to do one thing, a duplicate of a helper that already existed, a configuration knob wired to nothing, test-only surface left in a production module. The architecture and domain lenses read the system, so this reads as drift rather than as a hundred individually reasonable diffs — and a coherence family of rules covers the provable cases in C# today, each finding carrying the two spans that prove it. Those ship advisory: they inform the survey and never move the score until their false-positive profile is measured more widely. One sibling rule was cut for producing 71 false positives in 73.

Generated tests are the cheapest thing to produce and the easiest to fake. The test-quality read asks whether a test actually asserts, and reports skipped and assertion-free tests rather than counting files. This particular read is C# today — the language page states every such boundary rather than leaving you to find it.

A tax calculation that returns zero is the same finding in C#, Java, Python or Go, because the lens reads how the code is organised rather than its syntax. What differs by language is how much of the survey resolves — published per language, with the abstentions named.

Every survey scores the mix — the slop, fine and brilliant split — and trends it against the previous one. So the throughput argument and the quality argument become two readings taken the same day, instead of one conversation nobody can settle. Down is as useful as up: you find out within a sprint, not at the next incident.

An agent reporting success is not evidence. A resolve stays provisional until the next survey confirms it, and finding identity is line-insensitive, so a reformat cannot launder one away. The arbiter is the measurement — never the agent's own account of its work. That is what makes "we cleaned it up" a checkable claim rather than a report.