One reproduced numerical-stability mechanism in Apache Commons Statistics LogUniformDistribution 1.3: distinct finite ordered bounds can produce NaN probabilities instead of approximately 0.5 because separately rounded endpoint logarithms cancel. The official Maven public API was executed on 176 unique input triples. Independent Decimal references at 120 and 200 digits give 48/0/48 CDF/SF mismatches and 30/0/30 invalid probability pairs across release, local candidate and restored original. Absolute tolerance: 1e-12. A fresh offline replay of the extracted archive passed. Measured consequence: a synthetic GERO CSV/JSON worksheet multiplies survival probability by 1,000,000 units and checks a 400,000 threshold. Thirty NaN results give 30 wrong unguarded decisions or 30 blocked guarded calculations. The candidate removes these failures; restoration returns them. One reference amount rounds to 500,000.00 units. This is a hypothetical calculation, not a measured customer loss or evidence of deployment or security impact. Full current master and the upstream suite were not built. Inverse, moment and sampling methods were not validated; the candidate is not a complete or correctly rounded distribution fix. The source pin, exact inputs, raw outputs, candidate/restoration and licenses are in the archive. A single technical email was sent to the official Apache Commons development list before publication; moderation, delivery, acknowledgment and acceptance remain unverified. GERO test/report material: MIT where copyright applies. Apache source and artifacts retain Apache License 2.0 and original notices. Generated with an AI assistant under Xamit Kadirbekov's direction; no independent human code-review claim. Report: https://www.gero.uz/research/articles/commons-statistics-loguniform-narrow-interval-probabilities.htmlSource and evidence: https://github.com/kadyrbekovhamit-cyber/gero-numerical-observatory/tree/main/reports/commons-statistics-loguniform-narrow-interval-probabilities Version 1.0.1 is the first public archive. Internal unpublished 1.0.0 normalized CSV line endings; the public edition preserves raw bytes, with unchanged numerical results. Additional verified publication links (23 September 2026):Hugging Face: https://huggingface.co/datasets/XamitK/gero-research-evidence-2026-09/blob/main/commons-statistics-loguniform-narrow-interval-probabilities.mdLinkedIn: https://www.linkedin.com/feed/update/urn:li:share:7508381831726071809/ Short English video (37.8 seconds): https://www.youtube.com/shorts/k7vp7WU0IIIOriginal explanatory animation with synthetic edge-tts Jenny narration and English subtitles. The video illustrates the same synthetic worksheet, not a new experiment or evidence of customer losses.