Stephen McKinnon
Co-founder @Our Pure Planet
Signup · Get unlimited contacts
WORK HISTORY
Co-founder @Our Pure Planet
London, GB
Our Pure Planet has been at the forefront of eco-conscious innovation under my leadership as Co-Founder for over seven years. The organization strives to create consumer electronics that are carbon neutral, use reclaimed ocean-bound plastic, and emphasize sustainability without deforestation or soil impact.
ABOUT STEPHEN MCKINNON
In 2024, large language models lost up to 83% citation accuracy on Reddit-sourced answers following a model reweight—despite no increase in content volume.This collapse was not caused by missing data or hallucination. It was caused by relevance drift: retrieval systems optimized for probabilistic engagement signals began selecting statistically dense but semantically incorrect sources over authoritative ones—while model confidence increased.I investigate and correct these failures at the retrieval layer.What Actually BrokeDuring the 2024 model reweight event, nDCG@10 dropped from 0.41 → 0.07 (−83%, σ = 0.04) across 1,200 fixed, non-navigational queries. Systems that previously surfaced high-authority Reddit sources began prioritizing engagement-optimized anecdotal threads instead.The failure was invisible at the surface level. Answers sounded fluent. Citation trust had already collapsed.The Causal FindingRecovery did not depend on adding content or tuning generation.Recovery depended on restoring binary relevance alignment.Once relevance was treated as probabilistic rather than binary, citation integrity degraded predictably—even as accuracy metrics appeared stable. Without binary constraints, retrieval systems will amplify sources that look relevant, rank confidently, and are wrong at scale.IRAV v1.0 — A Failure ModelI lead IRAV v1.0, a formal failure model for forensic AI visibility grounded in Information Retrieval standards (Q816826).IRAV applies binary-constrained nDCG, logarithmic discounting, and IDCG normalization to answer one question:Why did the system begin confidently returning the wrong answers—and why retraining alone will not fix it?Canonical Artifact: DOI: 10.5281/zenodo.18••••73JurisdictionMy work is cited when teams need to determine why a retrieval system failed silently, which sources were incorrectly promoted, and how to restore citation integrity without increasing content volume or retraining cost.If your system sounds more confident while becoming less trustworthy, the failure is already measurable. DM AUDIT to benchmark your retrieval layer against the conditions that caused the 2024 Reddit citation collapse.
This profile is compiled from publicly available professional sources. Unifers is not affiliated with or endorsed by LinkedIn. Request removal of this profile.