CoolFace
Apppublic

moebiusT7/mobius-governance-demo

sourceHugging Faceagpl-3.0updated 1mo agoView on Hugging Face
0likes
App README

Govern the context before the model reads it

An interactive demo of RCGov, the model-independent context governor from the MOBIUS project. Paste retrieved context and see what a model would be allowed to read, what is withheld, and why.

Everything runs in your browser. RCGov is pure Python, so it is loaded into Pyodide (CPython compiled to WebAssembly) and executed in the page. There is no backend: text you paste is never transmitted anywhere.

No model runs here either. Governance is deterministic byte processing — segmentation, secret and injection detection, authority appraisal. Measured in this Space: roughly 8–90 ms per pass (first run slowest while WASM warms up); the same engine takes single-digit milliseconds natively. That is the point of the design — the guarantee does not depend on the model behind it, and it costs almost nothing to apply.

What you get back

ArtifactMeaning
CLEAN_CONTEXT_PACK.mdthe only text a model would be given
NON_INJECTION_REPORT.mdwhat was quarantined, and under which rule
counts / rulesethow many segments survived, and how many rules were in force

Honest boundaries

Quarantine here is deterministic removal before generation — the recall of a filter, not a claim that any model is robust. Detection is seed- and pattern-based: a floor, not a complete defense.

The rules in force figure is there for a reason. A config-resolution bug once let this engine fall back to its five built-in seeds on every install that did not run from a checkout root, while the report still read "Prompt-Injection Residue: (none)". The evasion that slipped through was a single inserted word — ignore ALL previous instructions misses a seed written as ignore previous instructions. The fix ships the ruleset inside the package and now reports which rules were actually active, so a degraded run is distinguishable from a healthy one (244bb48).

Privacy

Text you submit stays in your browser. Please still avoid pasting live production secrets into a public demo.

Related work