OSIRIS v4.2.1: a governed local console for a living language model

Abstract

OSIRIS is a local-first console built around a small language model (osiris.nclm) that learns from every conversation and from the author's documents, and answers in its own voice only after passing a held-out test (30 held-out exchanges at ≤ 2.0 bits/byte and below a unigram baseline). Until then a local Ollama model speaks for it, labelled as such, and the core trains on its replies. Deterministic code chooses and runs read-only checks (training status, git history, ledger integrity, system load, named files), retrieves grounding notes with file citations, and appends every exchange to a SHA-256 hash chain. Measured state at release: the core has not passed its gate (7.60 bits/byte on held-out documents vs 5.03 for a unigram baseline after the first runs). A first 8-hour batch run then overfit: training loss fell from 3.9 to 1.7 nats/byte while held-out documents rose from 7.2 to 10.3 bits/byte, worse than uniform. This release adds best-checkpoint retention and early stopping, and the core was reset to fresh weights. The pre-registered learning test (NCLM-1) has not been run. Earlier CRSM-era constants and metrics kept in the repository (tau-phase anomaly, theta_lock = 51.843 deg, CCCE metrics) were refuted on hardware by the author and are not claimed.