SlothGuard

SlothGuard is a prototype access reviewer built for the Hacktoberfest Open-Source AI Challenge, Week 3 ("Slow and Steady", 19-25 October 2026), whose brief asks for a project with open-source AI at its core that rewards laziness, and for an account of why open innovation matters. SlothGuard automates the dull part of an access review, finding segregation-of-duties (SoD) conflicts, and keeps a person in charge of the one decision that matters. Deterministic rules check 40 synthetic users against 9 SoD rules and produce 8 findings (5 high, 2 medium, 1 low). Gemma 4 E2B, an open-weight model run locally through Ollama, writes a short explanation for each finding; the code caps it at four sentences and falls back to a fixed template, and severity always comes from the rule, never from the model. A revocation is only proposed until a person approves it on a phone through Auth0 (Universal Login, CIBA, Guardian); a denial, a timeout (120 s) or any error means nothing is revoked, and every step is logged. On an ordinary PC without a dedicated GPU (Intel Core i5-1334U, 16 GB RAM), the explanation phase took 63.5 to 157.4 s across six recorded runs, including one with the PC offline, and turning off the model's thinking mode cut generated tokens by about 81%. The post states its limits openly: synthetic data, a small model that sometimes chooses the wrong wording, no comparison with a closed model, and no measurement of the time saved. Code is under the MIT License; documentation, figures, video and text are under CC BY 4.0.