MOBIUS Project

Study artifacts for 'Machine Doxa' — pre-registered consensus/differentiation study in LLM practice space: frozen prereg, norm/preference battery, raw sessions (7 models x 3 sessions), analysis, adversarial review record. Companion to machine-habitus.

Study artifacts for 'Measuring Machine Habitus' — pre-registered MCA of LLM disposition space: frozen prereg, 120-item forced-choice battery, raw sessions (7 models x 5 sessions), analysis code, adversarial review record.

Study artifacts for 'A Guardrail That Failed Quietly' (DOI 10.5281/zenodo.21903321): a silent prompt-injection ruleset degradation, its measured cost, and an adverse residual effect on the attacks the guardrail misses.

Mobius Reflective Context Governor — local-first semantic hygiene and context governance for LLM systems. Govern what a model is allowed to read before it answers.

QAT Q4_0 vs PTQ Q4_K_M under a fixed governance stack — a 2x2 quantization-regime study of Gemma 4 12B (dense) and a 26B-class MoE released pair, governed routing quality. Sequel to gemma4-mtp-quality-study.

Answer only when warranted; inject only what is fit to govern. MMV answer-entitlement + RCGov context governance as a model-independent layer + OpenAI-compatible proxy. AGPL-3.0; patent pending.

RSG-CFSI reference implementation — regularized financial stress indicator with out-of-sample discipline (companion code to MOBIUS Technical Companion Papers No.1). AGPL-3.0.

Cross-agent memory governance — gate what an AI agent's persistent memory may assert before the agent relies on it. RCGov's axiom (InjectContext => ContextReady) transplanted onto the memory layer of Claude Code / Codex.

Local-first, citation-grounded civic-RAG engine for 東京都議会 deliberation records — plenary + all committees, 1999–present (engine + facts only; ships no minutes text)

Memory-echo governance for LLM assistants with memory: provenance-tagged memory + deterministic output-time self-citation verification, with an empirical pilot. AGPL-3.0.