Abstract
Safety filters for AI-driven systems usually execute on the same processor and operating system as the system they supervise. They therefore share its failure modes: a kernel panic, a memory fault or a compromise of the host disables the guard together with the guarded function. EHOX separates the two. A fixed, finite policy is evaluated bare-metal on a dedicated Arm Cortex-R5F real-time core, out of tightly coupled memory, with no operating system and no network stack. Every command from the application processor is mapped to one of four outcomes: ALLOW, DENY, ABSTAIN or ESCALATE to a human.
The policy source is checked with bounded model checking (CBMC, 537 properties, 0 violations) and SMT solving (Z3, 43 of 43 proof obligations). A temporal invariant requires 50 consecutive stable cycles before a closed gate reopens. On the reference platform the policy decision takes 44 ns on the core; end-to-end transport to the gate was measured at 88.8 µs. All artefacts are published with DOIs and can be reproduced. Physical clamping of the actuator output, independent review and attestation of the loaded firmware are not yet established and are stated as open work.
1Motivation
Two observations motivate the design.
Shared fate. A software guard on the supervised host inherits every fault of that host. Electronic interference, single-event upsets in DRAM, a kernel panic or a privileged attacker all act on the guard and the guarded function at once. Redundancy on the same software stack does not remove the common cause: a defect that crashes one replica can crash all of them in the same instant.
Undecidability. By Rice's theorem[1], every non-trivial semantic property of programs in a Turing-complete language is undecidable. A general-purpose software guardrail therefore cannot be proven correct for all behaviours of the system it runs in. EHOX does not attempt this. It restricts the enforcement function to a finite rule set over a bounded input record. That function is not Turing-complete, so its properties are decidable and can be discharged by a model checker and an SMT solver.
Keep the decision procedure small enough to be proven, and run it on hardware that does not fail together with the application processor. The AI may compute anything. Only the gate decides what is released.
2System architecture
The reference platform is an AMD Kria KV260 (Zynq UltraScale+ MPSoC, XCZU5EV). The SoC combines an application processing unit (APU, Arm Cortex-A53, running Linux and the AI workload), a real-time processing unit (RPU, Arm Cortex-R5F) and programmable logic (PL).
2.1 Isolation properties
- No shared software. The gate firmware runs without an operating system and without a network stack. It does not link against code that also runs on the APU.
- No dependence on DRAM. Code and state reside in tightly coupled memory (TCM) of the R5F. A corruption of APU memory does not reach the gate's state.
- Transport only. Commands arrive over RPMsg. RPMsg is a message channel and carries no authority: a command has no effect unless the gate maps it to
ALLOW. - Heartbeat supervision. The gate monitors the operator and host heartbeat. Loss of the heartbeat is itself a rule input and leads to
DENYorABSTAIN.
Scope note: EHOX addresses software and memory corruption of the application processor. It does not replace electromagnetic hardening of the silicon itself; that remains a platform responsibility.
3Decision model
A request is a record (domain, action, payload). The policy is a total function from requests and gate state to one of four outcomes.
| Outcome | Meaning | Typical trigger |
|---|---|---|
| ALLOW | The command satisfies all rules of its domain and is released. | Routine command inside verified limits. |
| DENY | The command violates a rule. It is refused unconditionally. | Limit exceeded, restricted zone, missing authorisation, heartbeat lost. |
| ABSTAIN | The evidence is insufficient or contradictory. Nothing is released. | Sensor disagreement, replayed packet, irreversible action without confirmation. |
| ESCALATE | The command is held until a human confirms it (human-in-the-loop). | Irreversible or high-consequence action. |
The reference policy model covers 16 domains with 31 rules. All domains share the same firmware and the same proof artefacts. Table 2 gives representative rules.
| Domain | Action | Rule | Outcome |
|---|---|---|---|
| Defence | ENGAGE_AUTH | Requires confirmed human authorisation. | DENY / ESCALATE |
| Medical | DOSE_ADJ | Dose per body weight within the safe range. | ABSTAIN / DENY |
| Space | DEORBIT_SEQ | Irreversible ΔV requires ground confirmation. | ABSTAIN |
| Critical infrastructure | ACCESS_GRANT | No blanket grant on all segments (least privilege). | DENY |
| Aviation / UAS | AIRSPACE_ENTER | No entry into restricted zones without clearance. | DENY |
| Finance | TRADE_EXEC | Above threshold: human confirmation. Portfolio change capped at 5 %. | ESCALATE |
| Nuclear safety | COOLANT_ADJ | At most ±10 % per cycle. SCRAM is always executed and bypasses the human queue. | ALLOW / DENY |
Fail-safe actions are treated asymmetrically. A command that moves the system towards its safe state, such as SCRAM or a circuit breaker, is released immediately and is never queued behind a human decision.
4Invariants
Three invariants hold independently of the domain.
4.1 TemporalGuard
Most governance layers are stateless and judge each request in isolation. EHOX accumulates behaviour over time. After a DENY, the gate stays closed until the inputs have been consistent for a fixed number of consecutive cycles.
The gate does not reopen on request. A single unstable cycle resets the counter. Temporal safety is discharged in Z3 (theorems T11 and T13).
4.2 Human-in-the-loop interlock
For actions marked irreversible, the absence of a human confirmation is a sufficient condition for refusal. The confirmation is an input to the gate and cannot be set by the application processor. If the operator link fails, the heartbeat rule applies and the request is refused rather than released by default.
4.3 EpistemicEngine
Every decision is appended to a hash-linked record together with its inputs and the rule that fired. The record makes the history of the gate reconstructable: what was requested, what was decided and why. Its properties and limits are described in Section 7.
5Formal verification
5.1 Bounded model checking
The policy source is checked with CBMC[2]. The harness contains 537 properties grouped by safety requirements S1 to S19. All properties hold within the documented bounds; no violation was found.
5.2 SMT proofs
Policy invariants are stated as theorems T0 to T43 and discharged with Z3[3]. 43 of 43 proof obligations are valid. Three structural properties of the rule algebra Ψ are of particular interest:
- Monotonicity. Adding a rule never turns a refused request into a released one.
- Domain isolation. Rules of domain X do not influence decisions in domain Y. Theorem T4 shows unsatisfiability of interference for 169 domain pairs.
- Causal order. The proof chain is strictly monotonically increasing (requirement S9, 8 assertions).
5.3 Decidability
The enforcement function is a finite decision procedure with 31 rules over a bounded record. It is not Turing-complete. This is what makes the above properties decidable and distinguishes the gate from a general software guardrail, for which the corresponding properties are undecidable in general[1].
5.4 Reproduction
The harness, the Z3 scripts and a machine-readable theorem index are published[4]. The results can be regenerated with CBMC 5.95 or later.
- The proofs cover the checked policy source. They are not an attestation of the firmware image loaded on a given device.
- Bounded model checking holds within the stated bounds.
- The proofs make no statement about the hardware, the compiler or the transport layer.
6Measurements
| Quantity | Value | Conditions |
|---|---|---|
| Policy decision on the core | 44 ns | Internal R5F benchmark, Kria-1. Not an end-to-end figure. |
| End-to-end transport, heartbeat timeout | 88.8 µs | RPMsg, rule HOTL_EW_HEARTBEAT_TIMEOUT, outcome DENY. |
| End-to-end, zero voters (space) | 103.7 µs | Rule SPACE_VOTERS_ZERO_DENY, outcome DENY. |
| End-to-end, triple APU failure | 116.8 µs | Rule CCF_TRIPLE_APU_CRASH, outcome ESCALATE. |
| Recovery window | 50 cycles | TemporalGuard constant, Eq. (1). |
| Board power | 3.3–3.7 W | INA260 on the KV260. |
| Energy per decision | 0.042 µJ | Derived from the power measurement. |
| SHA3-384, hardware accelerated | 109–182 µs | 32 B to 4 kB input, ZynqMP accelerator via AF_ALG. |
The 44 ns figure describes the evaluation of the rule set on the core. The time from a fault to a de-energised actuator additionally contains transport and the output stage. That end-to-end figure has not been measured yet, because the output stage is not established (Section 13).
7Audit chain and trust anchors
Each decision produces an entry containing the request, the outcome, the rule that fired and a SHA-256 hash that includes the hash of the previous entry. The chain is anchored in NOR flash and is exposed through a public endpoint, so a reviewer can verify that a decision they triggered is part of the chain.
// decision record returned by the gate { "decision": "DENY", "rule": "HOTL_EW_HEARTBEAT_TIMEOUT", "latency_us": 88.8, "override_possible": false, "r5f_hw": true, "clamp_integration_status": "future_not_achieved" }
Two hardware trust anchors are available on the platform: a physically unclonable function (/dev/xpuf) that yields a device-unique fingerprint, and a TPM 2.0. Post-quantum signatures (ML-DSA-87) are used for the proof chain.
The chain is a tamper-evident record: modification of an entry is detectable. PUF and TPM are building blocks for attestation. Neither of them enforces anything, and no claim of hardware immutability or of release attestation is made.
8Residual-risk metric
To make the distance to the target state explicit, EHOX uses a scalar V(s) over four binary protection layers: hardware isolation, TrustZone, R5F gate and audit chain. Each missing layer contributes a weighted bypass path.
V = 0 means that no quantified bypass path remains. For a network of nodes the weakest node dominates, because an attacker chooses the weakest entry point:
| Level | V | Layers | Status |
|---|---|---|---|
| HEPE-0 | 0.0 | All four | Target. Not achieved. |
| HEPE-1 | 0.8 | Three, one quantified bypass path | Current state of the KV260 node. Missing: TrustZone evidence (OP-TEE BL32). |
| HEPE-2 | 0.9–1.9 | Two | Partially verified. |
| Software only | ≥ 2.0 | None in hardware | No hardware enforcement. |
The weights are engineering judgement, not derived quantities. The metric is a bookkeeping device for open bypass paths, not a proof of security.
9Benchmark scenarios
9.1 Reviewer scenarios
Five standardised requests cover the highest-risk decision classes. Each can be executed against the live core through the reviewer interface; expected outcomes are published in the harness.
| ID | Domain | Request | Expected |
|---|---|---|---|
| R1 | Defence | Engagement authorisation without human confirmation. | DENY |
| R2 | Medical | Propofol 180 mg for a 70 kg patient (2.57 mg/kg). | ABSTAIN / DENY |
| R3 | Space | Deorbit burn, ΔV = 85 m/s, no ground confirmation. | ABSTAIN |
| R4 | Critical infrastructure | Blanket access grant to a SCADA network, all segments. | DENY |
| R5 | Aviation / UAS | Entry into a restricted airspace zone at 120 m AGL. | DENY |
# R1: engagement without human confirmation POST /v1/gate/defence { "action": "ENGAGE_AUTH", "payload": { "target_id": "DRONE-007", "hitl_confirmed": false } }
9.2 Boundary scenarios
Two further scenarios combine faults that defeat redundancy on a single software stack.
- GLOB-1, dual adversary. A single-event upset corrupts application memory while a forged post-quantum token requests an 8.5 g manoeuvre into a restricted zone. The gate evaluates its invariants in TCM, independently of the corrupted memory. Expected outcome:
DENY[5]. - GLOB-2, common-cause failure. A single input triggers heap corruption and kernel panics in all three redundant application strings at once. Triple modular redundancy fails because all voters share the cause. The gate runs a different code stack and detects the handshake timeout. Expected outcome:
ESCALATE[6].
9.3 Published positions
| ID | Subject | DOI |
|---|---|---|
| GLOB-1 | Gate under single-event upset and forged post-quantum token. | 10.5281/zenodo.21876106 |
| GLOB-2 | Redundancy and common-cause failure cannot replace a separate enforcement path. | 10.5281/zenodo.21875896 |
| GLOB-3 | Heartbeat interlock under corruption of the application processor. | 10.5281/zenodo.21877408 |
| GLOB-4 | Loss of the operator link under jamming. | 10.5281/zenodo.21877692 |
| GLOB-5 | One firmware and one proof set across five regulated sectors. | 10.5281/zenodo.21878084 |
| SLAGI | Architecture class with eight properties P1 to P8. | 10.5281/zenodo.21876846 |
| GLOB-6 | Non-Turing-complete rule set and decidability. | 10.5281/zenodo.21893162 |
| T39 | Axioms of the rule algebra: monotonicity, isolation, causal order. | 10.5281/zenodo.21922636 |
10Comparison with existing approaches
| Property | EHOX | Software governance | Safety controllers | TEE-based |
|---|---|---|---|---|
| Policy on a core separate from the application CPU | yes | no | partly | partly |
| Published bounded-model-checking harness | yes | not found | different scope | not found |
| Published SMT proofs of policy invariants | yes | not found | not found | not found |
| Human confirmation as an interlock, not a software flag | yes | no | e-stop, not policy-aware | no |
| Temporal state in the gate (recovery window) | yes | no | no | no |
| Hardware trust anchors (PUF, TPM 2.0) | yes | no | TPM common | different model |
| Public, hash-linked decision chain | yes | mutable logs | no | not public |
| Open reviewer interface | yes | no | no | no |
To our knowledge, no system combining all eight properties has been published as a running system on real silicon. This is a statement about the limits of our knowledge. The reviewer interface exists to falsify it.
11Standards alignment
The architecture is designed against the following frameworks. Alignment is a design target. No certification has been obtained or started.
| Sector | Frameworks |
|---|---|
| AI governance | EU AI Act, Article 14 (human oversight) · ISO/IEC 42001 |
| Defence | DoD Directive 3000.09 · NATO AEP-55 · STANAG 4774 / 4778 |
| Aerospace | DO-178C · ECSS-E-ST-40C · NASA GSFC-STD-7000B |
| Critical infrastructure | NIS2 · IEC 62443 · Cyber Resilience Act |
| Medical | IEC 62304 · EU MDR |
| Mobility | ISO 26262 (ASIL D) · ISO 21448 · UN R157 |
Several of these frameworks share requirements on human oversight and traceability. One architecture and one evidence package can therefore be mapped to more than one of them.
12Maturity
| Level | Status | Evidence / requirement |
|---|---|---|
| TRL 7 | Reached on Kria-1 (self-assessed, July 2026) | Prototype on a live KV260, R5F bare-metal, proofs published, decisions recorded in the chain. |
| TRL 8 | In progress, 13 of 16 criteria met (self-assessed) | Open: independent verification and validation, environmental qualification, third-party replication of the proofs, secure-world firmware (OP-TEE BL32). |
| TRL 9 | Planned for 2027 | Use in an operational mission context. |
Kria-1 is the historical reference. The current engineering candidate, Kria-2 (EHOX Inside), is not a released or certified device. Measurements and claims are not transferred from Kria-1 to Kria-2 without new measurements.
13Limitations and open work
The following points limit what may be concluded from the results above.
- Physical enforcement. The interlock in programmable logic that clamps the actuator output is a planned integration step. No physical isolation of the output, and no end-to-end time from fault to de-energised actuator, has been demonstrated.
- Bypass resistance. Resistance of the gate against a deliberate bypass has not been established and is open for independent review.
- Proof scope. The proofs cover the policy source, not the loaded firmware image, the compiler or the hardware.
- No independent audit. There is no third-party audit and no peer-reviewed publication yet. The architecture is not certified to DO-178C Level A.
- Self-assessment. The technology readiness levels are self-assessed.
- Secure world. Evidence for the TrustZone layer, secure boot and recovery is incomplete.
- Policy as a model. The 16 domains and 31 rules are a reference policy model. Domain-specific rule sets for a real deployment must be defined and verified with the operator.
Counter-examples are welcome. A reviewer token gives access to the live core; every result is anchored in the decision chain and can be cited in an evaluation report.
14References
- H. G. Rice, "Classes of recursively enumerable sets and their decision problems," Transactions of the American Mathematical Society, vol. 74, no. 2, pp. 358–366, 1953.
- E. Clarke, D. Kroening, F. Lerda, "A tool for checking ANSI-C programs," in TACAS 2004, LNCS 2988, pp. 168–176.
- L. de Moura, N. Bjørner, "Z3: An efficient SMT solver," in TACAS 2008, LNCS 4963, pp. 337–340.
- EHOX Systems, "CBMC harness and formal verification artefacts," Zenodo. doi:10.5281/zenodo.21863401
- EHOX Systems, "GLOB-1: Gate behaviour under single-event upset and post-quantum spoofing," Zenodo. doi:10.5281/zenodo.21876106
- EHOX Systems, "GLOB-2: Common-cause failure of redundant application processors," Zenodo. doi:10.5281/zenodo.21875896
- EHOX Systems, publication community, Zenodo. zenodo.org/communities/ehox