Abstract
Techniques are described for routing stability in co-evolving decision systems that combine a machine-learning (ML) model and a large language model (LLM). A routing controller selects, per case, whether to use the ML tier, the LLM tier, or optionally human review. A pre-invocation economic gate permits an LLM call when expected economic value of improved decision quality exceeds LLM cost. After an LLM output is generated, a post-invocation epistemic gate calibrates the likelihood the output is wrong using a learned model conditioned on case features and evaluation signals, and selectively accepts or falls back. High-confidence accepted LLM outputs are used in a progressive distillation loop to retrain the ML model, reducing LLM workload. Routing accuracy is maintained under feedback-induced distribution shift using coordinated exploration and continuous recalibration based on outcomes from both exploration and production routing.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Routing Stability in Co-Evolving Machine Learning and Language Model Decision Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10609