Lost in Locale: Source Formats and Accepted Errors in a Local Browser Agent

Regional dates and amounts can be converted into syntactically valid but incorrect application values. This exploratory study evaluates one local Qwen3 4B agent with a model-generated table and deterministic calculator on 48 synthetic cases from six workflow families. Each case has regional and explicit source representations and three trials in separate text, baseline-browser and review-preview conditions. Of 864 planned executions, 862 ended; two user interruptions remain unobserved. Baseline browser runs produced 18 wrong accepted operations in 143 observed regional executions and 7 in 144 explicit executions. The primary paired, equally family-weighted difference was -7.14 percentage points, with an exploratory 95% interval [-20.14, +2.78] and family sign-flip p=0.25. Exact completions and absent writes are reported alongside errors. The review-stage preview supplied no observed reduction in wrong writes on complete pairs. The findings describe a fixed benchmark and one qualified policy; six related templates, serial collection and incomplete observations limit inference. They establish neither a population benefit nor equivalence. This record contains the preprint, an unchanged reproducibility archive with raw model/browser traces, frozen manifests, source history, analysis and the technical report, plus the current preprint source. Text answers are not application writes. The older protocol 0.1 pilot and Ollama 0.40.14 cohort are separate historical evidence and are not pooled. Status: exploratory preprint; not independently peer reviewed. No journal or conference acceptance is claimed. All transactions and entities are synthetic. Inference used local Ollama with no paid inference API. OpenAI Codex assisted software implementation, debugging, literature lookup and drafting; measured outputs came from the stated local model. MIT License.