I'm a 17-year-old independent researcher from Russia working on a new sequence model architecture. No university, no advisor, entirely self-funded. Harmonic is a 3-level hierarchical SSM with predictive coding — each level processes at a different timescale and receives prediction errors from the level below. O(L) compute and memory. Results on enwiki8 (28M params, equal token budget): +1.4% over Transformer at 1K tokens, +6.7% at 8K, +11.4% at 32K. At 64K both Mamba and Transformer OOM on H100 80GB — Harmonic trains successfully. Results replicate on WikiText-103. Preprint: https://doi.org/10.5281/zenodo.20381714 Logs: https://github.com/Omibranch/harmonic-logs.