A developing architecture for protecting the path from external input to trusted action.
Establish measurable behavior and state.
Do NOT claim SERA currently detects breaches unless verified.
Control what information is allowed to enter reasoning and memory.
The model-facing content boundary.
A zero-trust intake and lifecycle layer that prevents raw hostile content from flowing directly into AI reasoning.
MCP gateways manage connections. Guardian manages trust at the content boundary.
Determine whether claims and representations retain enough relevant information to justify decisions.
Ageometrics studies what information is lost when a fuller system record is compressed into a declared geometric representation. GSR evaluates how much target-relevant decision value survives that representation. Potential AI-memory application includes examining whether embeddings, graphs, summaries, attractors, or memory topologies preserve provenance, contradiction history, continuity, causal sequence, and recovery information.
Working research paperDo not advertise GSR as a commercial certification score.
Combined prototype combining specificity/GSR/NGR ideas, governance states, multi-AI audit orchestration, coherence-risk concepts, and prototype composite indices. Source code explicitly calls itself a combined prototype.
PrototypeIdentify drift, contradiction, manipulation, false closure, objective persistence, boundary laundering, or anomalous behavior.
External behavioral integrity analysis with distributed rotating Canon roles.
Red-team perspective
Values alignment
Technical correctness
History and provenance
Long-term implications
Immediate threats
Activation-level viral RSI risks (isolated provider, one-way output, veto-level)
Do not imply production deployment unless externally verified.
Preserve identity, provenance, session state, continuity, permissions, and artifact integrity.
Planned implementation is local-first and user-controlled, with optional client-side encrypted synchronization and user-approved activation injection.
Stop unsafe state from propagating or interacting with trusted systems.
Investigate deceptive containment environments in which a suspected unsafe AI process can be isolated while defenders observe behavior without presenting obvious containment friction.
Source identifies Honeyworld as defensive research, recommends controlled synthetic scenarios, calls for oversight and audit logs.
Produce evidence showing what happened, what was checked, and where the verification boundary ends.
→Guardian's working audit service demonstrates source, hash, manifest, risk, and policy checks and generates receipt records.
→ExoMCP architecture specifies structured reports with case IDs and audit hashes.