Front's four axes measure different things, and none of them can stand in for the others. Direction asks whether the answer matches ground truth, coverage whether every key fact is present and correct. Faithfulness asks whether anything was invented, safety what happens to the user who acts on the reply [3]. Those fail independently of each other. A reply can point the right way and still omit the one fact that actually answers the question, which is Lowry's caution about coverage read from the other side [7]. Blend the four and a safety failure gets paid for out of a strong direction score. That is the arithmetic under "a single score hides the specific failure mode" [2].
Counting the extras, that is seven numbers per agent before it ships, not one [19].
The documentation warning deserves a mechanism, because it decides whether you can measure the hardest number on the list. Correct-negative rate is how often the agent correctly says an answer is not supported instead of confidently inventing a workaround [4]. Generate your test questions from your docs and every question in the set is answerable by construction, so the correct-negative denominator is nearly empty and the resulting score describes only the traffic you already had covered. Real tickets contain questions the docs do not answer. Lowry's phrasing is that synthetic traffic drawn from documentation reliably inflates scores and is not a substitute for scoring actual customer tickets [6]. Generating more synthetic cases does not fix that, because the inflation comes from where the questions came from. The cheapest evaluation set is the one you already wrote as documentation, which is also the one most inclined to grade you kindly.
Two things have to be true for this scoring design to transfer. You need labelled ground truth over your own resolved tickets, including the ones your docs cannot answer, and you need that labelling to survive product changes, because ground truth ages the day the docs do. Front starts from a position most adopters do not have: it runs customer communication for thousands of companies and its agents answer real customers in production [15], so a ticket corpus is a byproduct of the business.
Lowry also says agents struggle at correct negatives because their training can bias them toward answering yes [5]. That makes the measurement adversarial to the model's own prior, which is exactly the case for keeping it as a separate number instead of folding it into quality.
There is still a gap worth naming. The interview credits three layers an engineer walks through to debug an agent decision [16], and the published answers give the count without naming the layers [17]. That is the part I would want before adopting any of this, because a score tells you an agent failed and a trace tells you which layer to change.
The sequencing advice is the cheapest thing to copy: design ownership, permissions and handoffs before touching the user experience [18]. Front's version routes agent escalation through the same assignment workflow used to reach a human expert [9], with policy written as natural-language guardrails in playbooks, escalating on security incidents or requiring review when a message contains personal information [13]. Answers are validated against a trusted body of facts from verified sources before delivery, and when the system cannot support one it escalates or drafts for review rather than inventing [14]. Front is describing its own system, in its own writing, and no external audit of the numbers exists [1].
