Auditing First-Finisher Inference: A Reproducible Replay Study
Year: 2026
Type: preprint
Abstract: A first-finisher policy changes the composition of a model pool as well as its response time. We adapt an established all-subsets estimator to audit that change using complete correctness and timing records. A local study fits 96 random forests with deliberately mixed tree budgets and records 103,680 single-image inference calls. Across three overlapping digits-data splits, eight-candidate serial-time replay reduces accuracy by 9.97–11.16 percentage points relative to a uniformly selected candidate. An exploratory decomposition attributes almost all of this loss to budget composition. A full independent timing rerun preserves the eight-candidate loss, but changes a 32-candidate result by 5.28 points. The artifact provides an inspectable stress test and sensitivity analysis, not a new estimator, an LLM result, or a concurrent-serving speedup measurement. AI assistance. AI tools assisted literature discovery, manuscript drafting, code generation and checking, mathematical checks, and reproducibility analysis. The supplied project records do not identify exact tool/model versions for this paper. The reported classifier predictions and serial timings come from executed computations; the supplement identifies separate synthetic simulations. No LLM inference or physical concurrent-serving experiment was performed. Materials and attribution. The classifier study uses the scikit-learn digits subset of Alpaydin and Kaynak (1998), Optical Recognition of Handwritten Digits, https://doi.org/10.24432/C50P49, CC BY 4.0. No digit images are included. The all-subsets estimator and concentration bound retain attribution to the cited prior work. The five-page supplement is included with the manuscript. Research draft, revision 3. Not peer reviewed. Release status. The uploader approved public release of this revision on 2026-11-08. Author-review and submission-approval notes in the PDFs were written during draft preparation. This is a preprint and has not been peer reviewed.
Authors: Ozdem, Musa Emin, Egri, Rumeysa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225597424