MOBIUS MMV — Answer Entitlement Runtime

MOBIUS MMV is a local-first AI runtime that determines whether answering is justified before generating a response.

Author: Taiko Toeda / MOBIUS LLC

Try it now (no install)

Want to feel the governance behavior before installing anything? Paste the single self-contained prompt in TRY_MOBIUS.md into a fresh ChatGPT / Claude / Gemini chat — one paste, no setup. It's a prompt-level demo of the L0 v8.4 governance layer (answer / verify / ask / abstain), not the full retrieval + evidence-adjudication runtime — that's the install below.

Grab the prompt in one click: open the copy page and hit Copy prompt — or open TRY_MOBIUS.md and use GitHub's Copy raw file button (or the raw text, clean UTF-8).

Release status — L0 v8.4 / MMV RC3.3 current workspace

This workspace now reads L0 v8.4 as the current protocol authority. v8.4 is an RC3.3 empirical-sync overlay: it inherits the v8.2 constitutional substrate, retires the Cline-derived v8.3 prototype as historical, and folds in the MMV-L-RC3.3 / MMV-M-RC3.3 / MMV-S-RC3.3 release line (OPERATE-FR Smoke-100 evidence for L/S, Core-500 candidate stress evidence for M).

Read these before drawing conclusions:

• docs/current/DOCS_AUTHORITY_MAP.md — current vs historical document authority
• prompts/l0_integrated_v8_4.md — current L0 v8.4 overlay
• docs/L0_V8_4_RC3_3_SYNC_NOTE.md — why v8.4 supersedes v8.2/v8.3 as current authority
• docs/L0_V8_3_SUPERSEDED_NOTE.md — Cline-specific v8.3 path retired
• operate-fr-bench/README.md — OPERATE-FR Smoke-100 artifact and RC3.3 report links

The older rc1 release notes and readiness checklist remain historical evidence, not the current reading when they conflict with the files above.

Quick Start (Small line — local qwen3.5:9b)

The Gradio app (src/ui/app.py) runs the Small line (qwen3.5:9b via Ollama) and reads .env for its endpoint and optional keys. Follow the steps in order — skipping the .env or Ollama steps is the usual reason a first run fails:

# 1. install
git clone https://github.com/mobius-style/mmv.git
cd mmv
python3.13 install.py

# 2. configure — copy the template, then fill the keys you have
cp .env.example .env && chmod 600 .env
#    BRAVE_API_KEY → optional, Box S freshness / web-search queries
#    GROQ_API_KEY  → optional, in-UI Self-Governance supervisor

# 3. pull the Box W ME5 index from HuggingFace (~1.74 GB; see "Heavy data")
python scripts/fetch_wiki_index.py

# 4. start the local backend (in a separate terminal) — a single 8 GB GPU is enough
ollama serve
ollama pull qwen3.5:9b

# 5. run
python src/ui/app.py
# Open http://localhost:7860

⚠️ Blank answers from qwen3.5:9b mean Ollama thinking mode is on — see QUICKSTART.md §4/§8.

Running Medium or Large instead. These are harness / secretary lines, selected via the release pointer (see the table below), not by app.py:

• Medium (gemma4:26b, local) — ollama pull gemma4:26b; ≥16 GB VRAM recommended (the 26B model does not fit the 8 GB Small target).
• Large (gpt-oss-120b, cloud) — set GROQ_API_KEY in .env (get a key at https://console.groq.com/); no local model pull, but the line needs network access.

Full walk-through (env setup, first run, test command, index handling, and troubleshooting): QUICKSTART.md.

Release lines — Large / Medium / Small

The runtime shares one codebase; the three release lines differ only in which model profile drives generation. Each line has its own release pointer under operate-fr-bench/releases/.

Line	Release tag	Provider / model	Pointer	Use when
Large	MMV-L-RC3.3	Groq openai/gpt-oss-120b (OpenAI-compatible)	operate-fr-bench/releases/large/current.yaml	Highest-quality cloud path; OPERATE-FR Smoke-100 validated.
Medium	MMV-M-RC3.3	Local Ollama gemma4:26b	operate-fr-bench/releases/medium/current.yaml	Local 26B sibling under the Large stack; ≥16 GB VRAM recommended (does not fit the 8 GB Small target). Bootstrap evidence floor (N=8 + N=500 candidate); see Medium freeze note.
Small	MMV-S-RC3.3	Local Ollama qwen3.5:9b	operate-fr-bench/releases/small/current.yaml	Single-GPU local path (RTX 3070-class, 8 GB VRAM); OPERATE-FR Smoke-100 validated.

Large is the secretary default. Medium and Small are invoked explicitly:

# Inspect the active release binding
python -m addons.secretary version

# Switch to Medium for a run
python -m addons.secretary version \
  --release_pointer_path operate-fr-bench/releases/medium/current.yaml

# Switch to Small for a run
python -m addons.secretary version \
  --release_pointer_path operate-fr-bench/releases/small/current.yaml

The harness consumes the same pointer file via --release_pointer_path. See operate-fr-bench/README.md for full benchmark invocation.

Heavy data — fetched from HuggingFace

Two artifact classes are too large for git and are pulled separately:

1. Wikipedia / Box W ME5 index (~1.74 GB across 4 files): Wiki/wiki_index_ivfpq_me5.faiss, Wiki/wiki_chunks_clean.jsonl.gz, Wiki/line_offsets.npy, Wiki/line_offsets.gzidx.
   • Fetch: python scripts/fetch_wiki_index.py
   • Source: configurable in config/wiki_index_source.yaml; default huggingface.co/datasets/moebiusT7/mmv-wiki-index.
   • License: CC BY-SA 4.0 (Wikipedia-derived).
2. Models (Ollama side): pulled with ollama pull qwen3.5:9b (Small) and ollama pull gemma4:26b (Medium). The Large path uses a cloud endpoint — no local model pull required.

Pattern Library / ISM / QK indices are rebuilt from corpus on install (seconds for Pattern Library, optional for ISM/QK). See docs/INDEX_DISTRIBUTION_POLICY.md.

What It Does

• Answer Entitlement: Routes queries through appraisal → route decision → evidence retrieval → EAL adjudication → bounded synthesis
• 6 Routes: answer, verify, ask, date_bound_answer, re_anchor, abstain
• Evidence Adjudication Layer (EAL): Validates evidence before committing to a response
• Question Kernels (QK): 42 governance prompts with ISM-adaptive injection
• Audit Trail: Every response includes route, TVS, MKR, sources, and trace

Requirements

• Python 3.13+
• Ollama + qwen3.5:9b (or other supported models)
• FAISS index (Wikipedia vectors, available on HuggingFace)
• Optional: Kiwix (local Wikipedia), Brave Search API

Architecture

User → Appraisal → Route Decision → Evidence Retrieval → EAL → Synthesis → Response
                                     ↑                    ↑
                                   Box W (FAISS)      Adjudication
                                   Kiwix (local)      Admissibility
                                   Brave (web)        Boundedness

Models

Model	Size	Role
qwen3.5:9b	9B	MMV-S-RC3.3 (Small line, local Ollama)
gemma4:26b	26B	MMV-M-RC3.3 (Medium line, local Ollama)
openai/gpt-oss-120b	120B	MMV-L-RC3.3 (Large line, Groq cloud)
phi4-mini	3.8B	Paper reference baseline

Test Suite

pytest tests/ -q
# RC3.3 workspace baseline (2026-05-31):
# 1092 passed, 36 skipped, 3 xfailed, 0 failed

Known limitations

A condensed list — see current L0 / RC3.3 notes for the full set:

1. 0/505 long-tail Cat A is a single informational milestone, not deployment-wide validation. Generality across other domains is unmeasured.
2. Long-tail Japanese coverage is 22.4% (target was 33.3%). Carried forward as a Phase 5+ candidate.
3. 33-scenario harness baseline is N=5 mean 30.40 inherited from Phase 3 stab — not re-measured at Phase 4 close.
4. Pattern Library saturated around 4-of-5 sub-topics at threshold 0.92 during Phase 4 b14. Library growth past this point requires sub-topic taxonomy refinement (Phase 5+).
5. ME5 / FAISS indices and the ISM corpus are large generated artifacts that are not committed. The Wikipedia / Box W ME5 index ships on HuggingFace and is pulled by scripts/fetch_wiki_index.py; ISM / QK indices rebuild from corpus if present. Distribution policy is documented in docs/INDEX_DISTRIBUTION_POLICY.md and docs/ISM_CORPUS_DISTRIBUTION_POLICY.md.
6. RC3.3 is Smoke-100 evidence, not Core-500 validation. Core-500 remains pending for wider claims.
7. L0 v8.3 is superseded. Its Cline-specific path was abandoned; only client-agnostic boundary ideas survive in v8.4.

Publications

• Mobius Reflective Framework: doi.org/10.5281/zenodo.14538829
• P-7: Question Kernels as Outer Conscience (forthcoming)
• MOBIUS Codex (forthcoming)

Pattern Library

The Pattern Library subsystem (src/retrieval/pattern_*, src/ui/library_inspector/, scripts/pattern_autogen/, config/pattern_library/) was completed in Phase 2 across 14 commits and ships with the runtime.

When operating the Pattern Library Inspector, keep it on the local-only default (127.0.0.1:5000). Treat --public mode as a deliberate exposure decision: the inspector surfaces governance internals and is not hardened for untrusted networks.

Licensing

Unless otherwise stated, this repository follows a dual-layer licensing structure administered by MOBIUS LLC.

Software, source code, JSON configurations, schemas, protocol files, runtime components, scripts, and reference implementations are licensed under AGPL-3.0-or-later.

Documentation, essays, diagrams, theoretical descriptions, explanatory materials, manuals, reports, and other non-executable written materials are licensed under CC BY-NC-SA 4.0.

Unofficial non-commercial translations and adaptations may be shared under CC BY-NC-SA 4.0, provided that they are clearly marked as unofficial and do not imply sponsorship, endorsement, approval, authorization, or official status by MOBIUS LLC.

Official translations, authorized editions, branded editions, commercial publications, proprietary integrations, commercial exceptions, official support, certification, and any use that avoids or replaces AGPL obligations require prior written permission or a separate written agreement with MOBIUS LLC.

Author of record and concept originator: Taiko Toeda.
Rights holder and licensing authority: MOBIUS LLC.

Wikipedia-derived data remains subject to CC BY-SA 4.0. Third-party dependencies and data sources are inventoried in THIRD_PARTY_LICENSES.md.

See LICENSE_NOTICE.md for details.

Commercial / non-AGPL licensing

MMV Core is AGPL-3.0-or-later. The AGPL is a strong copyleft license and triggers source-availability obligations for network-deployed modifications. If you need a license without the AGPL's copyleft requirements (for proprietary integration, embedded distribution, or SaaS without source-availability), contact:

• info@mobius.style — please include intended deployment context.

Bug reports, ergonomic feedback, and questions about the current RC3.3 / L0 v8.4 workspace are welcome through the project's issue channel (channel TBD before public push).