AWS's turn-level metric separates the one broken turn from the three that inherited it

The failure class that makes this argument work is a successful tool call with a wrong argument. In the AWS example the agent selects the right action for the sales report and passes "profit" where "revenue" was expected ^[3]. Nothing raises. The trace looks clean, the artifact renders, and the later turns go on refining something correctly shaped and wrong.

So the scoring definition matters more than the block diagram. AWS defines truthfulness as factual consistency between the values the agent produced and what was expected, covering parameter values inside tool calls as well as statements in prose ^[6]. To compute that at turn 2, the judge needs the expected parameter value for turn 2. That is a different label from the one an outcome score needs, and an eval set built around final answers does not contain it. Completeness is a cheaper check by comparison, because it asks whether required elements are present rather than whether they are right ^[6]. Underneath both, AWS puts tool and action selection as the structural foundation rather than a third sub-metric ^[7].

Here is the arithmetic that makes the case, in the currency of turns. The example runs five turns, the error starts at turn 2, and it propagates through turns 3, 4 and 5 ^[4]. Four turns therefore emit wrong output and exactly one of them needs a change, so three of the four bad turns, 75 percent, are inherited ^[14]. That ratio is the whole argument for per-turn attribution, and it comes from a hand-built illustration rather than a measured trace population.

The composition step is where I would push. AWS describes sub-metrics measured independently and then recombined into a top-level score ^[16], and in this material the recombination is named rather than numbered. That gap matters because the post's own complaint against goal completion is that a score of 70 percent does not tell you whether the failures were factual errors, missing information, or wrong tool choices ^[11]. A composite whose weights you cannot inspect reproduces that opacity one level up. The diagnostic value sits in the per-turn sub-metrics; the single indicator on top is for a dashboard.

What AWS claims as new is the decomposition into named, separately measurable sub-metrics tracked turn by turn, which it says existing tooling does not provide ^[10], and it says so plainly: the headline contribution is the decomposition itself ^[8]. In my context that is the right trade when trajectories are long and the tools take parameters, because the thing you end up fixing is one parameter mapping at one turn rather than a conversation. On a two-turn retrieval bot, the outcome score already points at the only turn that could have broken, and the per-turn labels are pure overhead.