APPLIED SCIENCE & TECHNOLOGY · 6.1 / 6.2 / 6.3

Four Research Thrusts in High-Assurance Autonomous Execution Control

Investigating mathematical, cryptographic, and systems-engineering guarantees that bound autonomous software and cyber-physical agents when prompts, weights, or intermediate hosts are contested or compromised.

PLANE 01 · UNTRUSTEDAI / Agent Workload• mTLS + SPIFFE /agent• Zero target API keys• Zero signing authority• Zero direct egressTyped IntentPLANE 02 · GOVERNANCEController & OPA Gate• Default-deny Rego policy• Entropy & canary guard• RATS TPM2 / SEV-SNP• Runtime tensor check2-of-3 ReservePLANE 03 · QUORUMEd25519 Witnesses• Single-use (nonce, id)• Survives 1 node loss• Anti-rollback quorum• Monotonic hash anchorBound GrantPLANE 04 · ACTUATORTarget Interlock• Key-bound verification• <9ms P95 stop barrier• Signed effect receipt• DDIL hold (no blind retry)
THRUST 01 · APPLIED RESEARCH & PROTOTYPING (6.2 / 6.3)

Cryptographically Separated Execution Authority

Eliminating ambient authority and in-process guardrails across autonomous software, C2 decision aids, and automated industrial workflows.

Scientific Problem

When an autonomous agent shares memory space, API tokens, or signing keys with its policy evaluator, any prompt injection, deserialization flaw, or supply-chain compromise collapses governance into full actuator privilege.

Technical Approach

  • Workload isolation with mTLS + SPIFFE identity and zero direct network egress.
  • Out-of-process default-deny OPA policy evaluation over canonicalized typed intents.
  • Target-side verification requiring cryptographic capability grants bound to exact action digests.
THRUST 02 · APPLIED RESEARCH & PROTOTYPING (6.2 / 6.3)

DDIL Partition-Tolerant Quorum & Non-Atomic Reconciliation

Preventing split-brain double actuation, VM snapshot replay, and blind retry storms across denied, degraded, intermittent, and limited-bandwidth (DDIL) links.

Scientific Problem

Tactical RF, acoustic, and contested SATCOM links routinely drop acknowledgement frames after physical or external side effects have already occurred. Naive retry loops duplicate irreversible actions, while restored VM snapshots can replay stale nonces.

Technical Approach

  • 2-of-3 Ed25519 witness quorum enforcing single-use reservation semantics across independent state stores.
  • Survives single witness loss or simultaneous controller + single-witness rollback with zero replay commits.
  • Explicit UNKNOWN_UNTIL_SIGNED_RECEIPT state machine holding conflicting retries until signed target effect receipts reconcile.
THRUST 03 · APPLIED RESEARCH (6.2)

Hardware-Rooted Tactical Attestation (IETF RATS)

Binding runtime authorization grants to verifiable platform, enclave, and model-artifact measurements without cloud endorsement reachability.

Scientific Problem

Standard cloud confidential-computing workflows assume continuous internet access to vendor endorsement certificate authorities and ignore post-boot mutation of model weights or adapter tensors in memory.

Technical Approach

  • Disconnected-enclave verification using pre-provisioned endorsement bundles for TPM 2.0 and AMD SEV-SNP evidence.
  • Freshness-bound attestation claims coupled directly into policy and witness grant issuance.
  • Keyed runtime integrity verification over loaded model weight and adapter segments.
THRUST 04 · FOUNDATIONAL & APPLIED RESEARCH (6.1 / 6.2)

Dual-Space Model Integrity, Covert-Channel Bounds & Delegated Subagents

Bounding latent adapter backdoors, high-entropy covert exfiltration, and privilege amplification across multi-agent hierarchies.

Scientific Problem

Fine-tuned open-weight models and delegated subagent workflows introduce dormant trigger surfaces, steganographic outbound channels in structured tool calls, and confused-deputy escalation across child tasks.

Technical Approach

  • Pre-deployment structural screening of low-rank adapters paired with runtime telemetry guards.
  • Deterministic outbound schema canonicalization, canary traps, and Shannon-entropy anomaly gates.
  • Monotonic capability attenuation (A_child <= A_parent) and sub-10ms cascading tree revocation.