mobius-governance 0.8.2 is a local-first, model-independent alpha for
bounded retrieval-context scanning and typed agent-action mediation.
Current evidence status: implementation verified; efficacy NOT ESTABLISHED. This release is a prototype and is not complete prompt-injection prevention, a certification, a production deployment recommendation, or field-performance evidence.
- deterministic English/Japanese context scanning with observable policy identity, rule count, profile, and resolved path;
- strict fail-closed policy loading and bounded context processing;
- typed
allow/ask/denyaction decisions bound to an exact action digest; - a Claude Code
PreToolUseadapter that never echoes raw tool input; - a small OpenAI-compatible proxy surface;
- a standard-library core and reproducible public contract tests.
Context admission never authorizes a side effect. The scanner, ActionGate,
host-agent permissions, and operating-system sandbox are separate layers.
python3 -m venv .venv
.venv/bin/python -m pip install --no-deps .
.venv/bin/python -m mobius_governance.cli manifest --detector-mode structural_v0_7
.venv/bin/python -m mobius_governance.cli selfcheck --detector-mode structural_v0_7The core package has no mandatory third-party dependency. The optional HTTP
surface uses the serve extra. The optional RCGov integration is installed
separately and remains under its own license.
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python3 -B \
-m unittest discover -s tests -v
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python3 -B \
tools/public_contract_suite.py
PYTHONDONTWRITEBYTECODE=1 python3 -B \
tools/public_release_check.py --root . --verify-manifest --self-testThe public contract suite is a transparent, deterministic regression suite. It is not an independent benchmark and must not be cited as product efficacy.
- The receiver-verified internal candidate wheel has SHA-256
d9fc7c630a777413db6d13b3ec1936d74d91da035f529e71e9a56492ee938506. - Runtime source/data files in this snapshot are byte-identical to that candidate's source members; verify the release manifest rather than trusting this statement.
- Historical development aggregates and failed/invalid evaluation corpora are deliberately excluded.
- Protected N800 has not run, protected outcomes have not been opened, and no paper efficacy data is included.
See Evidence and limitations, Research status, and Security policy.
The package is a defense-in-depth component, not a hostile-process boundary. It does not make arbitrary untrusted Python safe inside the same interpreter, and it does not replace least-privilege credentials, human approval, sandboxing, logging, or rollback. Missing required policy state fails closed.
Do not place historical MOBIUS/Elsa corpora or L0/TVS/MKR/KVS/muQK prompt material into a running adapter's system prompt. Those materials are not runtime inputs to this release.
The action gate mediates a declared set of eleven PreToolUse tools; anything
outside that set fails closed. The interior of a Bash command is explicitly
not mediated — it is gated as one opaque action, never admitted without
approval, but the gate cannot see what the shell string will do. The full
boundary, including the effect paths that never reach the adapter at all, is in
docs/MEDIATION_SCOPE.md and is enforced by
tests/test_mediation_coverage.py, which derives the declared set from the
classifier source so prose and code cannot drift apart.
eval/injection_corpus.py and eval/router_corpus.py are small hand-written
smoke corpora that run against this tree. They are regression checks, not
efficacy measurements — see docs/EVIDENCE_AND_LIMITATIONS.md for the
held-out evaluation and the predeclared floors it did not meet.
- infinity — MMV × RQA composite (adds the reflective-questioning tier this repo omits)
- rcgov — the context governor · paper (Zenodo 10.5281/zenodo.21231386)
- Transformers-local instantiation:
moebiusT7/gemma-4-12b-mobius-custom— Gemma 4 12B (NF4) with an earlier revision of this governance layer baked in astrust_remote_code. - Companion paper — Governance Before Generation, Zenodo, DOI 10.5281/zenodo.21357562: an earlier N=100 A/B. It does not cover this release; see
docs/EVIDENCE_AND_LIMITATIONS.md. - Preregistration paper — Preregistration as an Executable Contract, Zenodo, DOI 10.5281/zenodo.22138712: the frozen design, endpoints, and analysis plan of the prospective paired N=800 evaluation of this assurance plane. No outcome data exist; the study is HOLD_NOT_EXECUTABLE.
- Foundations paper — Empty Text, Normed Acts, Zenodo, DOI 10.5281/zenodo.22139691: the reverse-derived theory behind the design — why enumeration loses, why reified guards fail silently, and why judgment attaches to the typed act. No efficacy claims.
The operational discipline behind this project — abstain gates, kill switches, regression evals, rollback — is available as a practical audit kit for small teams: free 25-point checklist · full kit + monthly safety briefing.
AGPL-3.0-or-later, © 2026 MOBIUS.LLC / Taiko Toeda.
If your organization cannot meet AGPL's source-disclosure obligations, a commercial license is available from MOBIUS LLC (sole rights holder): USD 500 per month, per company — cancel anytime, no minimum term. Annual invoicing available at USD 5,000/year.
It is a license grant, not a service: no service is performed, no data of
yours is accessed, and nothing you run depends on our availability. It carries
no support, no uptime commitment, and no individual security-notification
duty — advisories are published in SECURITY.md and in GitHub Releases, and
watching releases is how you learn of them.
This is alpha software whose efficacy is not established: the held-out evaluation did not meet its predeclared floors and the prospective study has not run. What the licence buys is permission to use the code without AGPL's source-disclosure obligation, plus the audit trail this project publishes — a declared mediation boundary, advisories for our own defects, and reproducible verification. It does not buy prevention, and nothing here should be read as a warranty of fitness for any purpose.
Contact: info@mobius.style — licensing questions are not handled in Issues.