In layaAgent, a 421M encoder that cannot write chooses tools and arguments by multiple choice

"NONE carries a lot of the design," the author wrote ^[7]. Laya can pick a wrong span, but it can only choose from the spans the extractor pulled out ^[5]. For "How tall is that iron tower in Paris?", the right argument is "Eiffel Tower", and that string appears nowhere in the goal ^[6]. Laya's only honest pick is NONE. The step then waits for System 2, which the author describes as "the only component allowed to write a new value" ^[6].

The weak spot is the span-picking itself. Fixed-list arguments score 90%, above the 71% overall. So the other arguments, the ones picked from extracted spans, must score below 71% ^[11].

The gate decides which of those picks run unattended. A step runs automatically only if Laya is sure ^[13]. The obvious signal is the model's own confidence, 1 minus normalized entropy. The author found it badly compressed on this checkpoint: a 12-option question reads 0.1 to 0.3 even when Laya is right ^[13]. A threshold on that range cannot separate correct picks from wrong ones. The available text ends partway through that section. It stops before describing the replacement signal, and before the task-suite comparison behind the title's claim that the 421M encoder beat a 1.5B LLM at running the agent ^[18].

The approval guard does not rely on that gate. The author's complaint about the usual big-model loop is that nothing in its structure stops it from turning a request to delete the user's notes into an action ^[17]. In layaAgent, System 1 also asks once whether the goal changes something ^[15]. A miscalibrated confidence score cannot turn that check off.

Auditability is the claim I find best supported. Every Laya decision is a probability for each option, produced in one deterministic forward pass ^[3]. Its input is bounded too: the goal, the last four steps at one line each, and the last observation, all under a fixed token budget ^[19]. In the live run of "What's 18% of 2,450?", Laya picked the calculator at p 0.99, took the span from the goal and stopped, with no generative model involved ^[16]. The worked example is arithmetic, the one task where I would have skipped the LLM anyway.