Abstract
A method is disclosed for adding a stochastic explorer to a deterministic language-model routing policy without re-implementing that policy's precedence. A selection stage first invokes the unchanged deterministic rule chain with per-worker usage history withheld, so the history-threshold rules being replaced cannot fire. The chain's terminal no-signal rule returns an exported sentinel reason string. The stage samples only when the returned reason equals that string exactly, and otherwise returns the chain's answer unchanged. On the sentinel, a Beta-Bernoulli Thompson sampler draws one value per distinct routable candidate from that candidate's arm posterior, and routes to the highest draw, with ties going to the cheaper tier. An error or invalid value while forming candidates or sampling returns the chain's own keep. A scheduled job recomputes the posterior counts in one transaction from recorded run outcomes. It credits each run to the model named by the latest model-bearing execution-side record correlated with that run and started no later than the run ended, never to the router's decision, skips runs that lack such a record, and does not count as a failure a failed run whose recorded disposition carries a designated exclusion mark. The job does not require the sampler to be enabled, because the sampler reads the counts only when a master switch, a stage switch and a sticky hash-bucketed per-worker rollout all admit it.
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Sentinel-Gated Thompson Sampler Subordinate to a Deterministic Model-Routing Precedence Chain, with Execution-Record Run-to-Model Attribution and an Inert-Until-Gated Arm Refresh", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/13201