The ordering in the graph is the load-bearing part. Deterministic code runs first, and the SLM tone evaluator is invoked only if that check passes [11]. A draft with a wrong total therefore never receives a tone score, so fluent prose cannot buy its way through on style. Numbers get settled in Python, in integer cents, matched against ledger transactions [10].
$1,990 against $1,975 is a $15 overstatement, 0.76 percent of the invoice [16]. As hallucinations go that is almost courteous, and it is also exactly the size that survives a review by eyeball. The imaginary transaction ID TX-9999 stands out immediately [4], but the fifteen dollars sitting inside a polite paragraph is the harder catch.
When the check fails, the error goes back to the drafter as a structured critique under a retry budget of N <= 2 [12]. That means one request gets at most three passes through the model before the graph stops trying and escalates to a person [17][14]. Running locally, the API bill for those retries is zero [1], so the budget is priced in wall clock rather than dollars.
The Abstain branch is the one most linear agent designs leave out. If a client has zero unpaid transactions, or the records are ambiguous, the agent refuses to draft at all [13]. That answers the failure mode the author names directly: default prompting pushes a model to guess rather than cleanly refusing or escalating [5]. Refusal has to exist as a state in the machine itself.
The evaluation layer gets the same treatment. The trap described in the write-up is teams eyeballing five outputs, or asking another LLM to rate a response from 1 to 5, without measuring statistical rater agreement or testing for length bias [6]. Quadratic Weighted Kappa is how this design answers that, calibrating the judges instead of trusting their scores [7]. Approved and edited drafts then feed continuous memory, rather than the corrections disappearing into an email client the way they do without the loop [14][15].
What would have to be true for this to move to another workload: the quantity in dispute must be computable from records you already hold. Invoicing qualifies, because a balance is checkable against ledger transactions and line items are checkable against each other [9]. Where no such function exists, the deterministic branch has nothing to assert, and the gate collapses back to a language model judging a language model, which is the thing the architecture set out to distrust [8].
One honest limit on the evidence. What the author published is a blueprint with failure modes and code patterns, not a measurement report: there is no before-and-after error rate and no reported QWK value [18]. The design reasoning holds on its own terms. The cost per approved draft is a number an adopter will have to produce themselves.
