Local-first AI tooling and pre-registered research from Tokyo. We build systems that decide whether an AI should answer or act before it does — and publish the measurements, including the failures.
| Repository | What it is |
|---|---|
| mmv | Local-first conversational runtime with Answer Entitlement: decides whether answering is justified before responding |
| mmv-voice | Local-first speech-to-text: Whisper + on-device Gemma punctuation; words and numbers never change |
| book-ocr-studio | Turn books, PDFs and scans into searchable local text |
| mobius-governance | Answer-entitlement + context governance as a model-independent layer and OpenAI-compatible proxy |
| rcgov | Govern what a model is allowed to read before it answers |
| memgov | Gate what an agent's persistent memory may assert before the agent relies on it |
| mobius-browser-guard | Auditable mediation layer for browser-acting agents, with its own failure map |
| mmv-secretary | Self-evolving secretary addon with a reversibility charter and auto-rollback |
More: infinity · rqa · tokyo-insight · rsg-cfsi
Pre-registered studies with frozen protocols, raw sessions and adversarial-review records: machine-habitus · machine-doxa · reflective-budget-governor · answer-entitlement-prereg · guardrail-silent-degradation · gemma4-quant-regime-study · gemma4-mtp-quality-study
Sponsor MOBIUS on GitHub — $5, $25 or $250 a month. Sponsorship funds maintenance and continued open release. It is not a license and includes no support entitlement. Licensing is set per repository; for AGPL-licensed code, a separate commercial license is available — info@mobius.style.
Supporters are listed (optionally) in SUPPORTERS.md.