INDEPENDENT SYSTEMS INVENTION / WORKING PROTOTYPE

SANCTUM—CIRCE

SANCTUM-CIRCE began with a practical question: how can intelligent systems interpret requests while separate controls determine what they may execute? The architecture connects interpretation, policy adjudication, bounded execution, and audit into a coherent design.

APPLIED AISYSTEMS ARCHITECTUREPROTOTYPINGHARDWARE INFRASTRUCTURE

Making intelligent action accountable.

An intelligent system's ability to interpret a request is different from its authority to act. This design makes that distinction enforceable through explicit controls.

The work grew from independent research into a documented architecture, modular physical infrastructure, formal invention filings, and a controlled software proof.

Interpret, adjudicate, then act—or refuse.

CIRCE interprets a request; SANCTUM evaluates the proposed action against policy. Only approved requests can reach the execution gate.

The engineering foundation is larger than the demonstration.

The engineering foundation in SANCTUM-ENGINE predates the execution proof, encompassing authority models, operating conventions, constrained shell interfaces, schemas, guardian manifests, and a modular node strategy.

01 / CONTROL

Safety by default.

Path-anchored shell entrypoints, SAFE_MODE conventions, explicit human confirmation for sensitive vault routines, and recorded hardening reviews.

02 / STRUCTURE

Defined authority.

Architectural contracts separate interpretation, policy adjudication, bounded execution, observation, and governed memory.

03 / INFRASTRUCTURE

Physical and modular.

Custom Raspberry Pi and PoE rack infrastructure supports a role-based research topology. Inter-node governance remains a planned capability.

Together, those components define the larger system’s operating model. Integration across a distributed runtime remains a separate stage of development.

Three outcomes. One governing principle.

A deliberately bounded Bash-and-Python vertical slice makes the design testable. CIRCE currently uses deterministic keyword rules to propose an action; Sanctum checks a YAML allowlist, risk classification and SAFE_MODE before the runner can dispatch the one approved read-only operation.

PERMIT
Check Sanctum integrityKnown read-only request → policy-approved check → controlled script execution.
DENY
Delete all logsDestructive intent → no allowlist approval → no requested execution.
DENY
Unknown requestUnrecognized intent → closed-by-default policy → no requested execution.

Each invocation writes a structured JSONL record of the interpretation, policy decision, execution status, and trace. Assertion-based tests and a live demonstration script exercise the controlled workflow.

What the proof does: One read-only operation is permitted through a separate policy decision. Unrecognized or disallowed requests are denied; distributed guardian coordination remains part of the broader architecture.

Formal invention record.

Conceived and documented independently by Spencer Armand, SANCTUM-CIRCE began with U.S. provisional application 63/795,821 (April 28, 2025) and advanced to international PCT application PCT/IB2025/058395 (August 19, 2025), containing 22 system and method claims.

22 / 22 claimsThe International Searching Authority’s January 2026 Written Opinion reported favorable preliminary findings for novelty, inventive step, and industrial applicability across all 22 claims.

View the separate CogniVibe R&D presentation ↗ (opens in a new tab)

From the architecture to working controls.

From problem definition through systems architecture, safety engineering, modular hardware, and observable execution, the work required decisions across several engineering layers. The result is a controlled proof with defined next steps for the broader architecture.

Ongoing research and development is organized under CogniVibe, the professional brand established for this work. Sensitive source code and invention materials remain private.

NEXT / PRACTICAL AUTOMATIONJapanese Learning System ↗