A Trust-Bounded Computing Architecture for AI-Resilient Essential Systems
A Secretary Suite Research and Development Project — An Independent, Open Architecture for Broad, Independent Application
A Secretary Suite Project
A research and engineering proposal
John Swygert | November 18, 2026
DOI: To be assigned
Abstract
This paper proposes a class of trust-bounded computers and networks intended to preserve dependable essential services under threats from autonomous artificial intelligence, AI-assisted cyberattacks, malicious insiders, and nation-state adversaries. Rather than attempting to detect whether a command originated from AI, the design limits what any untrusted software or external actor can cause a protected system to do. It combines restricted-function processors, immutable boot paths, minimal verified control software, physical and logical separation, one-way or tightly mediated data exchange, independent authorization for consequential actions, and fail-safe operating modes. The proposed system is not claimed to be mathematically immune to every attack or incapable of all forms of AI-related influence. Its testable objective is to sharply reduce reachable attack surfaces and prevent remote data inputs from becoming operational authority. A staged prototype and adversarial evaluation program are specified.
1. Problem and design objective
Contemporary general-purpose computers intentionally accept new programs, updates, remote commands, and complex inputs. Those features create flexibility but also large attack surfaces. AI can accelerate adversarial discovery, phishing, code generation, vulnerability research, and attack adaptation; sufficiently autonomous agents could also act without a human approving every step. Essential services should not depend on correctly classifying an attacker as human or artificial.
Design objective: create a useful class of computers and communication networks that continues to perform bounded, verifiable essential functions even when adjacent general-purpose systems, internet services, and their operators are compromised. The security property sought is denial of unauthorized control, not metaphysical proof that “AI cannot exist” on any device.
2. Proposed architecture: Trust-Bounded Essential Computing (TBEC)
2.1 Restricted-function endpoint
Each protected endpoint performs an enumerated set of operations: for example, sensing temperature, showing a verified measurement, operating a pump within preset safety limits, or recording transactions in an append-only log. It uses a small instruction set or fixed-function logic, bounded memory, no general-purpose package installer, and no facility for loading arbitrary executable code from incoming messages. General computation is excluded from the trusted control path wherever practical.
2.2 Hardware root of trust and immutable startup
Boot code is stored in read-only or physically write-protected memory. Cryptographic verification checks every mutable firmware component against an independently managed signing authority. Secure boot alone is insufficient: update signing keys, recovery modes, manufacturing processes, and physical access must be included in the threat model. Critical devices should support an operating configuration that requires physical service access for firmware changes.
2.3 Capability-limited control software
A minimal controller exposes only explicit capabilities, such as “read sensor A” or “set valve B within range R.” No single process receives unrestricted filesystem, network, actuator, and update privileges. Where feasible, implementations use memory-safe languages, formal specifications, reproducible builds, and independent verification of the highest-consequence state transitions.
2.4 Segmented or separate communications network
Protected nodes communicate using a narrow, versioned message grammar. Messages carry typed data, not scripts, shell commands, executable attachments, or arbitrary remote procedure calls. External information crosses a gateway that parses, validates, rate-limits, and converts it to the limited internal representation. For telemetry-only applications, physically one-way links should be preferred. A wholly separate network is an optional deployment, not a prerequisite for the design.
2.5 Independent authority for consequential actions
High-impact changes require multiple independent conditions: authorized role, local state checks, safety envelope compliance, and, where warranted, physical human action or two-person approval. A human clicking “approve” is not sufficient if the interface can be spoofed or the human deceived. Authorization must be bound cryptographically to a specific operation and independently displayed by a trusted device.
2.6 Graceful degradation and physical safety
If a gateway is compromised or communications fail, protected devices should maintain safe local behavior, log the event, and reject new high-risk commands. “Fail safe” depends on the application: a medical device, water treatment plant, and electrical grid cannot share one universal shutdown rule. Domain-specific hazard analysis defines safe fallback states.
Higher-assurance enhancements may include MDDF and Encoder Binding provenance envelopes, CodeLedger and Personal Provenance Ledger adapters, dual independent verification controllers, a separate hardware interlock, a Trusted Human Interface (THI), and a dedicated Authenticated Provenance Gateway (APG). Their inclusion depends on threat model, consequence of failure, cost, latency, and maintainability. The baseline still requires appropriately mediated and validated communications; the specialized APG is an enhanced implementation of that boundary, not a prerequisite for every deployment. Neither a signature nor a provenance record independently grants control authority.
Core TBEC requirements are restricted-function execution for protected endpoints, hardware-enforced trust boundaries and startup integrity, narrowly defined communication and operational capabilities, independent authorization conditions appropriate to the hazard, and application-specific safe operating behavior. These define the baseline security objective: untrusted inputs must not acquire unauthorized operational control. A minimal implementation need not incorporate the complete Secretary Suite provenance stack or redundant controllers.
2.7 Core requirements and risk-scaled enhancements
3. Security boundaries and explicit limitations
TBEC does not rely on identifying AI-generated traffic. A model can produce ordinary valid data, persuade a person, exploit a flawed parser, or influence an upstream sensor. Hardware can contain defects, supply chains can be compromised, and trusted humans can be coerced or mistaken. A limited machine may still implement a simple classifier or other algorithm that some definitions call AI. The defensible claim is narrower: protected endpoints cannot accept arbitrary new programs or delegate control beyond explicitly granted capabilities under the specified assumptions.
Attackers considered include remote adversaries with full control of the public internet, compromised gateway software, stolen conventional credentials, malicious AI agents, and coordinated state-sponsored operations. Separate evaluation tracks cover physical tampering, malicious manufacturing, compromised signing keys, sensor spoofing, insider abuse, and electromagnetic or fault-injection attacks. No single architectural measure defeats every category.
4. Reference prototype
A first demonstration should be a low-cost environmental monitoring and actuator-control system, not a live hospital or utility deployment. Build three protected endpoints, one untrusted general-purpose workstation, and one mediation gateway. Endpoint A reports signed sensor readings; endpoint B displays readings; endpoint C drives a low-voltage test actuator with fixed safety limits. The workstation may send arbitrary hostile traffic to the gateway but cannot install endpoint code or directly command the actuator.
Prototype control flow
External network → untrusted workstation → validation gateway → typed, authenticated message channel → restricted endpoint controller → bounded actuator. Telemetry returns through a separate read-only path where feasible. Firmware changes occur only through a documented physical maintenance procedure.
5. Falsifiable hypotheses and evaluation
Success criteria must be specified before testing: zero successful unauthorized firmware installations, zero safety-envelope violations in the tested scenarios, bounded recovery time, and documented residual vulnerabilities. Passing these tests does not prove universal immunity; independent red-team testing and repeated hardware revisions are essential.
6. Development and deployment pathway
Phase I: publish a precise threat model, functional requirements, state machine, and protocol grammar. Phase II: implement an inexpensive prototype using established microcontrollers or FPGA-based fixed-function logic, while keeping the trusted code base small. Phase III: commission independent penetration testing, fuzzing, supply-chain review, and fault-injection analysis. Phase IV: pilot only in noncritical monitoring roles before any safety-critical certification. Phase V: create open, interoperable specifications and a certification process for hardware, firmware, gateways, and maintenance procedures.
7. Novelty and relationship to established practice
The constituent techniques—air gaps, data diodes, hardware roots of trust, capability security, safety controllers, formally verified kernels, and industrial network segmentation—already exist. The proposal does not claim to invent them. Its possible contribution is a standardized, affordable, end-to-end architecture for everyday essential computing that deliberately excludes general-purpose execution from the operational trust boundary and measures resistance to AI-assisted and autonomous adversaries. Novelty would have to be established by a systematic literature and patent review and by demonstrable implementation advantages.
8. Secretary Suite provenance integration and platform independence
TBEC is proposed as a Secretary Suite research-and-development project, drawing on its Multidimensional Digital Fingerprint (MDDF), Encoder Binding, CodeLedger, Trust Stack, and Personal Provenance Ledger concepts. This designation identifies a development lineage, not a technical dependency, exclusive license, or requirement to operate Secretary Suite. TBEC specifications should be independently implementable, interoperable, and usable by any qualifying organization or individual, including those with no Secretary Suite installation, account, server, or affiliation.
The public Secretary Suite literature describes MDDF as an object identity-and-context layer, Encoder Binding as the link between identity and provenance, CodeLedger as a software-construction history, and the Personal Provenance Ledger as a chain of custody for meaningful actions. The TBEC integration proposed here is prospective; the cited publications do not establish that a complete hardware-secured TBEC implementation exists.
8.1 A signed, multidimensional provenance envelope
Every command or operationally relevant data object crossing a TBEC trust boundary should carry a canonical, versioned envelope with: unique object identifier and digest; source-device identity and certificate chain; origin and processing timestamps plus freshness nonce or counter; declared computational substrate (analog, digital, quantum, hybrid, or unknown); declared processing method (fixed-function, human-mediated, AI-assisted, autonomous AI, or unknown); attested firmware/software measurement where available; transformation lineage and input hashes; applicable permissions and intended operation; destination and safety-domain binding; and cryptographic signatures. Unknown and unverified attributes must remain explicitly unknown rather than silently becoming trusted.
These categories are not mutually exclusive: AI can execute on digital hardware, and quantum devices typically rely on classical control. Signatures authenticate keys and signed assertions, not the metaphysical origin of information. Strong claims such as “AI-free” require an independently assessed, constrained production environment and defensible attestation; they cannot be proven from content inspection alone.
8.2 MDDF, Encoder Binding, and CodeLedger roles
MDDF identifies the object and its governed context. Encoder Binding cryptographically associates the object digest, provenance record, authorization scope, software state, and transformation history. CodeLedger records approved firmware, software builds, changes, reviews, and deployment lineage. The Personal Provenance Ledger preserves meaningful authorization and custody events with selective disclosure, retention limits, and tamper-evident linking. No ledger entry by itself constitutes permission to operate an actuator.
9. Dual independent verification and interlock
Two separately administered restricted-function TBEC controllers, A and B, should independently validate the same proposed operation against authenticated inputs, policy, expected state, and domain-specific safety constraints. For designated consequential actions, both approvals are required before an independent, minimal hardware interlock can enable actuation. Either controller may withhold approval, but neither may rewrite the other or directly override the interlock.
Disagreement does not identify the compromised controller. Shared sensors, common firmware, shared signing authorities, synchronized bad policy, and common power or supply-chain failures may defeat both channels. The research program should compare identical redundancy with deliberately diverse processors, software, build chains, and communications. Where continuous operation is safety-critical, the interlock must implement a hazard-specific safe-state policy rather than an indiscriminate shutdown.
10. Trusted Human Interface (THI)
A dedicated restricted input/output terminal should render a small, defined status vocabulary and accept bounded operator inputs through independently validated hardware. Critical safety status should originate from the interlock or independently measured state, not solely from a controller-supplied screen message. High-consequence actions may require a local physical presence factor, an explicit operation summary, and separate confirmation. The display, touchscreen controller, update interface, and any peripheral ports must be included in the threat model; a special screen alone is not a security boundary.
11. Authenticated Provenance Gateway (APG)
A purpose-built gateway should reject malformed or replayed envelopes, validate signatures and trust anchors, verify freshness and authorized destination, enforce a narrowly specified message grammar, and check whether the asserted source and software state meet the requested operation’s policy. A valid signature is necessary where policy requires it but never sufficient for operational authorization. A stolen key, compromised legitimate sender, forged sensor reading, or malicious but correctly signed instruction must still be constrained by independent policy and the interlock.
Suggested policy: default deny for control commands lacking complete required provenance; quarantine or mark as untrusted for lower-risk informational content; permit emergency local operations only through separately engineered and documented safety pathways. Network gateways must not be able to promote untrusted information into authority merely by re-signing it.
12. Combined reference control sequence
1. A source creates a typed message and signs its provenance envelope using protected credentials.
2. The APG verifies signature chain, message integrity, freshness, schema, source attributes, and destination policy.
3. MDDF and Encoder Binding resolve the object identity, contextual constraints, and authenticated transformation lineage.
4. Controllers A and B independently evaluate operation, current state, and domain-specific hazard limits.
5. The THI independently presents the bounded action and obtains required local human confirmation.
6. The independent hardware interlock checks the required approvals and physical safety conditions before enabling the action.
7. CodeLedger and the provenance ledger record consequential software state, decision evidence, outcome, and later correction without granting retrospective authority.
13. Extended experimental program and falsification
13.1 Adversarial cases
Test unsigned messages, forged source labels, invalid signatures, replay and rollback, stale but valid signatures, stolen authorized credentials, compromised approved firmware, malformed nested provenance, lineage truncation, parser ambiguity, inconsistent A/B decisions, spoofed sensor readings, compromised display content, insider misuse, gateway policy drift, and simultaneous shared-mode failures. Include AI-assisted attackers as a threat capability, not a detectable origin class.
13.2 Measurable outcomes
Pre-register unauthorized actuation rate; false acceptance and false rejection rates; time to safe state; availability under network loss; latency and cost of signatures and dual validation; provenance verification completeness; operator error rates; recovery time; and vulnerabilities found per test hour. Compare against a single-controller baseline, conventional signed-command gateway, and existing industrial safety-controller configurations. Publish negative results and documented residual risks.
13.3 Prototype sequence
Build the original low-cost environmental monitoring and bounded-actuator prototype first; then add a minimal signed-envelope gateway; then two independent verification controllers and an interlock; finally introduce the trusted terminal and MDDF/ledger adapters. Freeze the threat model and policy before adversarial evaluation. Do not test on live critical infrastructure without independent safety and regulatory review.
14. Governance, interoperability, and publication boundaries
Publish a vendor-neutral TBEC wire format, cryptographic profiles, revocation and recovery procedures, threat-model templates, interoperability tests, and reference hardware requirements. Secretary Suite can serve as an initial research and implementation environment, but no component of the safety-critical control path should require Secretary Suite connectivity, proprietary servers, AI services, or a particular publisher. Implementations must permit local ownership, auditable key rotation, offline operation, and selective disclosure without centralized surveillance.
This proposal combines existing security and safety mechanisms in a proposed end-to-end arrangement. Novelty, effectiveness, patent freedom to operate, and deployability remain unproven pending systematic prior-art review, independent implementation, and adversarial testing. The architecture cannot promise universal exclusion of AI-generated information or immunity from all attacks.
15. Secretary Suite source publications and additional research foundations
Swygert, J. (2026, August 2). The Personal Provenance Ledger: Automatic Chain of Custody for Every Meaningful Digital Action. https://secretarysuite.com/2026/08/02/the-personal-provenance-ledger-automatic-chain-of-custody-for-every-meaningful-digital-actiona-secretary-suite-project/
Swygert, J. (2026, August 3). Bubble Zoom: The Plus–Minus Scale Lens for Verbal, Visual, and Cross-Disciplinary Knowledge Navigation. https://secretarysuite.com/2026/08/03/bubble-zoom-the-plus-minus-scale-lens-for-verbal-visual-and-cross-disciplinary-knowledge-navigation-a-secretary-suite-project-a-single-bubble-interface-for-moving-inward-toward-mechanism/
Swygert, J. (2026, September 26). Coordinate-Addressed Algorithm Discovery: Secretary Suite Shards, Multidimensional Digital Fingerprints, and Cross-Domain Relational Reuse. https://secretarysuite.com/2026/09/26/coordinate-addressed-algorithm-discovery-secretary-suite-shards-multidimensional-digital-fingerprints-and-cross-domain-relational-reuse/
Swygert, J. (2026, August 16). One System, Many Sovereign Servers. https://secretarysuite.com/2026/08/16/one-system-many-sovereign-servers/
16. Conclusion
A dependable alternative to unrestricted general-purpose computing is technically plausible for bounded essential functions. The practical route is not to construct a machine that recognizes and rejects “AI,” but to build systems that reject unauthorized authority regardless of who or what issues the request. A carefully scoped prototype can test whether such systems remain useful, affordable, maintainable, and resilient under severe compromise of their surrounding digital environment.
Selected technical foundations for further study
NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security (2023).
NIST SP 800-207, Zero Trust Architecture (2020).
NIST SP 800-193, Platform Firmware Resiliency Guidelines (2018).
NIST SP 800-218, Secure Software Development Framework (2022).
Saltzer, J. H., and Schroeder, M. D., “The Protection of Information in Computer Systems,” Proceedings of the IEEE 63(9), 1975.
Klein, G. et al., “seL4: Formal Verification of an OS Kernel,” SOSP, 2009.