Novel Odor Intelligence (NOI) v0.5: Family-Level Open-Set Recognition, Reliability-Aware Memory, and Repeat-Sensing Decisions on Physical-Sensor Data

Neuro-Olfactive Intelligence (NOI) v0.5 — Technical Report This technical report presents a locked computational evaluation of the Neuro-Olfactive Intelligence (NOI) research program using recorded physical-sensor data from SmellNet-Base. The evaluation employed five deterministic odor-family rotation folds, non-overlapping 128-sample windows, 66 extracted features, and six sensor channels. It investigated family-level novel-odor rejection, repeat-sensing decisions, and reliability-aware memory under a predefined evaluation protocol. Key Findings The sentinel method reduced the false-known rate for unknown odor families from 0.802057 under the confidence baseline to 0.572281. However, its known-family acceptance rate was 0.760019, below the prespecified minimum of 0.80. The asymmetric-veto variant achieved a false-known rate of 0.503333 and a known-family acceptance rate of 0.722842. The sentinel unknown-detection AUROC was 0.682235. Reliability-aware memory produced a reported top-1 performance change of +0.083939. Under the registered evaluation criteria: H10 — Not supported: The required known-family acceptance condition was not met. H11 — Not supported: The asymmetric-veto strategy did not improve bounded decision utility. H12 — Supported: The reliability-aware memory comparison passed its two registered checks. These findings demonstrate a tradeoff between reducing unsupported known-odor classifications and maintaining acceptance of known odor families. They do not establish definitive chemical identification, human-equivalent olfaction, safety-critical performance, or deployment readiness. Conceptual Hardware Illustration The accompanying GuardianX NOI hardware illustration depicts a proposed electronic-nose architecture intended for future research and engineering development. It is conceptual only. No physical GuardianX NOI prototype was constructed, integrated, or experimentally validated as part of this study. Research Transparency and Reproducibility This release includes the technical manuscript, computational evaluation results, per-window evidence, and the validation-lock configuration supporting the reported findings. The original SmellNet sensor recordings are not redistributed. Research Organization: GuardianX LLC, Washington, United States. Related Research: NOI v0.3 — https://doi.org/10.5281/zenodo.22180610 Dataset Reference: Feng, D., Dai, W., Li, C., Pernigo, A., Wen, Y., and Liang, P. P. (2026). SmellNet: A Large-Scale Dataset for Real-World Smell Recognition. ICLR 2026. https://arxiv.org/abs/2506.00239