Action-Set-Informed Capacity Search and Output-Form Comparison for Compact Language Models in Smart-Lighting Control

Chat-to-action systems map variable requests to a finite set of executable device commands; can enumerating that set inform compact-model design? The SentiLight policy yields 240 command tuples from 1.3 × 1011 actuator combinations. This count and exposure per action frame a multi-seed search over width, depth, tokeniser and budget, without a demonstrated saving in runs. Under a fixed step budget, the width needed for a given accuracy rises with action count across four nested policies, and three random-subset draws keeping every attribute category reproduce the loss on shared actions at d = 384. A 15.7 M-parameter configuration attains 86.5% exact match on development data and 58.7% on unseen frames; its difference from a retrained 23.4 M reference is uncertain (+1.2 points; 95% interval −2.1 to +4.6). On that backbone, a 240-class head is 28 times faster than a generative decoder; their ranking reverses on post hoc sets. The decoder, the existing service adapter, runs as a 17.1 MB mixed-bit export at 95 ms median CPU latency behind a catalogue gate in an Android–FastAPI–Tasmota installation. All evaluation sets are constructed: an author-audited request set (24.9%) and a machine-paraphrase stress test (7.9%) locate the input-language range without bounding spontaneous requests.