What problem was I trying to solve?
Low-budget hyperparameter optimization in mixed search spaces is brittle. Categorical choices, noisy early measurements and sparse trial budgets make it easy for a proposal mechanism to overreact to a candidate that is not actually reliable.
What did I build?
A reliability-controlled selection layer. The language model is used as a semantic proposal engine only; what constrains the final choice is downstream of it — candidate validation, calibration-aware scoring, reliability-coupled risk control, and a reliability-gated episodic memory. The release is scoped to the manuscript: configs, scripts, figures, and minimal aggregate artifacts.
What actually happened?
Two experiments on CIFAR-100 carry the main evidence. A third is included as a failure-mode stress test, and the repository states plainly that it is a diagnostic rather than cross-domain proof. The manuscript is under review; citation metadata will be updated when a public preprint or final version exists.
What did I learn?
At a low budget, constraining the selector buys more than strengthening the proposer. And scoping a release to exactly the claims a manuscript makes — then naming that boundary in the repository — is what keeps a reproducibility claim defensible.
What can be verified?
Python 3.10.19, CUDA 11.8, PyTorch 2.4.1+cu118, with a Dockerfile and compose target and a recorded runtime report. Released results include aggregate metrics, seed-level convergence CSVs, checksummed manifests and the manuscript figure PDFs; an integrity script validates the delivery against the manifest. There is no paper link yet, because there is not yet a public one.